diff --git a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
index 095a833..8f66b8a 100644
--- a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
@@ -30,7 +30,11 @@
 {
 class ITensor;
 
-/** Interface for the dequantization layer kernel. */
+/** Interface for the dequantization layer kernel.
+ *
+ * @note The implementation supports only 3D input tensors
+ *
+ */
 class NEDequantizationLayerKernel : public INEKernel
 {
 public:
@@ -48,12 +52,12 @@
     ~NEDequantizationLayerKernel() = default;
     /** Set input, output, min and max.
      *
-     * @param[in]  input  Source tensor. Data types supported: U8.
-     * @param[out] output Destination tensor. Data types supported: F32.
-     * @param[in]  min    Minimum value of the input tensor.
-     * @param[in]  max    Maximum value of the input tensor.
+     * @param[in]  input   Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data type supported: U8.
+     * @param[out] output  Destination tensor with the same dimensions of input. Data type supported: F32.
+     * @param[in]  min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor.
+     *                     The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32
      */
-    void configure(const ITensor *input, ITensor *output, const float *min, const float *max);
+    void configure(const ITensor *input, ITensor *output, const ITensor *min_max);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
@@ -61,8 +65,7 @@
 private:
     const ITensor *_input;
     ITensor       *_output;
-    const float   *_min;
-    const float   *_max;
+    const ITensor *_min_max;
 };
 }
 #endif /*__ARM_COMPUTE_NEDEQUANTIZATIONLAYERKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h b/arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h
new file mode 100644
index 0000000..5e01acf
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h
@@ -0,0 +1,77 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef __ARM_COMPUTE_NEMINMAXLAYERKERNEL_H__
+#define __ARM_COMPUTE_NEMINMAXLAYERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+#include <cstdint>
+#include <mutex>
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the kernel to perform min max search on a 3D tensor. */
+class NEMinMaxLayerKernel : public INEKernel
+{
+public:
+    /** Default constructor */
+    NEMinMaxLayerKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEMinMaxLayerKernel(const NEMinMaxLayerKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEMinMaxLayerKernel &operator=(const NEMinMaxLayerKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEMinMaxLayerKernel(NEMinMaxLayerKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEMinMaxLayerKernel &operator=(NEMinMaxLayerKernel &&) = default;
+    /** Default destructor */
+    ~NEMinMaxLayerKernel() = default;
+
+    /** Initialise the kernel's input and outputs.
+     *
+     * @note output[0] = minimum
+     * @note output[1] = maximum
+     *
+     * @param[in]  input  Input tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data type supported: F32.
+     * @param[out] output Output tensor with shape [2, batches, ...] which stores the minimum and maximum value for each 3D input tensor.
+     *                    The dimensions over the second must match the batched dimensions of the input tensor. Data types supported: F32
+     */
+    void configure(const ITensor *input, ITensor *output);
+    /** Resets global minimum and maximum. */
+    void reset();
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+    void update_min_max(float *out_ptr, float min, float max);
+    const ITensor *_input;
+    ITensor       *_output;
+    std::mutex     _mtx;
+};
+}
+#endif /* __ARM_COMPUTE_NEMINMAXLAYERKERNEL_H__ */
\ No newline at end of file
diff --git a/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h
index 92cd142..617a2da 100644
--- a/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h
@@ -30,7 +30,11 @@
 {
 class ITensor;
 
-/** Interface for the quantization layer kernel. */
+/** Interface for the quantization layer kernel.
+ *
+ * @note The implementation supports only 3D input tensors
+ *
+ */
 class NEQuantizationLayerKernel : public INEKernel
 {
 public:
@@ -48,12 +52,12 @@
     ~NEQuantizationLayerKernel() = default;
     /** Set the input, output, min and max.
      *
-     * @param[in]  input  Source tensor. Data types supported: F32.
-     * @param[out] output Destination tensor. Data types supported: U8.
-     * @param[in]  min    Pointer to the minimum value of the input tensor.
-     * @param[in]  max    Pointer to the maximum value of the input tensor.
+     * @param[in]  input   Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F32.
+     * @param[out] output  Destination tensor with the same dimensions of input. Data types supported: U8.
+     * @param[in]  min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor.
+     *                     The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32
      */
-    void configure(const ITensor *input, ITensor *output, const float *min, const float *max);
+    void configure(const ITensor *input, ITensor *output, const ITensor *min_max);
 
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
@@ -61,8 +65,7 @@
 private:
     const ITensor *_input;
     ITensor       *_output;
-    const float   *_min;
-    const float   *_max;
+    const ITensor *_min_max;
 };
 }
 #endif /*__ARM_COMPUTE_NEQUANTIZATIONLAYERKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h b/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
index 7cd8360..8985861 100644
--- a/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
@@ -27,7 +27,6 @@
 #include "arm_compute/runtime/IFunction.h"
 
 #include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
-#include "arm_compute/runtime/Tensor.h"
 
 #include "arm_compute/core/Types.h"
 
@@ -37,6 +36,8 @@
 
 /** Basic function to simulate a dequantization layer. This function calls the following NEON kernels:
  *
+ * @note The implementation supports only 3D input tensors
+ *
  * -# @ref NEDequantizationLayerKernel
  *
  */
@@ -47,12 +48,12 @@
     NEDequantizationLayer();
     /** Configure the kernel.
      *
-     * @param[in]  input  Source tensor. Data types supported: U8.
-     * @param[out] output Destination tensor. Data types supported: F32.
-     * @param[in]  min    Minimum value of the input tensor.
-     * @param[in]  max    Maximum value of the input tensor.
+     * @param[in]  input   Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: U8.
+     * @param[out] output  Destination tensor with the same dimensions of input. Data type supported: F32.
+     * @param[in]  min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor.
+     *                     The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32
      */
-    void configure(const ITensor *input, ITensor *output, const float *min, const float *max);
+    void configure(const ITensor *input, ITensor *output, const ITensor *min_max);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h b/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
index ab189fe..d91b4ad 100644
--- a/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
@@ -26,7 +26,7 @@
 
 #include "arm_compute/runtime/IFunction.h"
 
-#include "arm_compute/core/NEON/kernels/NEMinMaxLocationKernel.h"
+#include "arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h"
 #include "arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h"
 #include "arm_compute/runtime/Tensor.h"
 
@@ -38,7 +38,9 @@
 
 /** Basic function to simulate a quantization layer. This function calls the following NEON kernels:
  *
- * -# @ref NEMinMaxKernel
+ * @note The implementation supports only 3D input tensors
+ *
+ * -# @ref NEMinMaxLayerKernel
  * -# @ref NEQuantizationLayerKernel
  *
  */
@@ -49,8 +51,8 @@
     NEQuantizationLayer();
     /** Set the input and output tensors.
      *
-     * @param[in]  input  Source tensor. Data types supported: F32
-     * @param[out] output Destination tensor. Data types supported: U8
+     * @param[in]  input  Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F32
+     * @param[out] output Destination tensor with the same dimensions of input. Data types supported: U8
      */
     void configure(const ITensor *input, ITensor *output);
 
@@ -59,9 +61,8 @@
 
 private:
     NEQuantizationLayerKernel _quantize_kernel;
-    NEMinMaxKernel            _min_max_kernel;
-    float                     _min;
-    float                     _max;
+    NEMinMaxLayerKernel       _min_max_kernel;
+    Tensor                    _min_max;
 };
 }
 #endif /* __ARM_COMPUTE_NEQUANTIZATIONLAYER_H__ */
diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
index 3bf2b35..70984f0 100644
--- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
@@ -23,9 +23,9 @@
  */
 #include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
@@ -35,16 +35,16 @@
 using namespace arm_compute;
 
 NEDequantizationLayerKernel::NEDequantizationLayerKernel()
-    : _input(nullptr), _output(nullptr), _min(nullptr), _max(nullptr)
+    : _input(nullptr), _output(nullptr), _min_max(nullptr)
 {
 }
 
-void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output, const float *min, const float *max)
+void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output, const ITensor *min_max)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_ERROR_ON_NULLPTR(min);
-    ARM_COMPUTE_ERROR_ON_NULLPTR(max);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(min_max);
+    ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() < 3);
 
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, DataType::F32, 0);
@@ -52,17 +52,20 @@
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
 
-    _input  = input;
-    _output = output;
-    _min    = min;
-    _max    = max;
+    _input   = input;
+    _output  = output;
+    _min_max = min_max;
 
     constexpr unsigned int num_elems_processed_per_iteration = 8;
 
     // Configure window
     Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-    update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), output_access);
+    AccessWindowStatic     min_max_access(min_max->info(), 0, 0, 2, min_max->info()->dimension(1));
+
+    // Update window and padding
+    update_window_and_padding(win, input_access, output_access, min_max_access);
     output_access.set_valid_region(win, input->info()->valid_region());
 
     INEKernel::configure(win);
@@ -74,31 +77,55 @@
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    Iterator input(_input, window);
-    Iterator output(_output, window);
+    Window window_input_output(window);
+    window_input_output.collapse_if_possible(INEKernel::window(), 3);
+    window_input_output.set(3, Window::Dimension(0, 1, 1));
 
-    const float32x4_t vmin    = vdupq_n_f32(*_min);
-    const float       range   = *_max - *_min;
-    const float32x4_t scaling = vdupq_n_f32(range / 255.0f);
+    Window window_min_max;
+    window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape());
+    window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1));
+    window_min_max.collapse_if_possible(INEKernel::window(), 1);
 
-    // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255]
-    execute_window_loop(window, [&](const Coordinates & id)
+    Iterator input(_input, window_input_output);
+    Iterator output(_output, window_input_output);
+    Iterator min_max(_min_max, window_min_max);
+
+    execute_window_loop(window_min_max, [&](const Coordinates & id_batch)
     {
-        const uint8x8_t  val_u8       = vld1_u8(reinterpret_cast<uint8_t *>(input.ptr()));
-        const uint16x8_t val_u16      = vmovl_u8(val_u8);
-        const uint32x4_t val_u32_low  = vmovl_u16(vget_low_u16(val_u16));
-        const uint32x4_t val_u32_high = vmovl_u16(vget_high_u16(val_u16));
-        float32x4_t      val_low      = vcvtq_f32_u32(val_u32_low);
-        float32x4_t      val_high     = vcvtq_f32_u32(val_u32_high);
+        // Get the min and max
+        const float min = *(reinterpret_cast<const float *>(min_max.ptr()) + 0);
+        const float max = *(reinterpret_cast<const float *>(min_max.ptr()) + 1);
 
-        // Dequantize -> (q / 255.0 * range) + min
-        val_low  = vmulq_f32(val_low, scaling);
-        val_high = vmulq_f32(val_high, scaling);
-        val_low  = vaddq_f32(val_low, vmin);
-        val_high = vaddq_f32(val_high, vmin);
+        const float32x4_t vmin    = vdupq_n_f32(min);
+        const float       range   = max - min;
+        const float32x4_t scaling = vdupq_n_f32(range / 255.0f);
 
-        const float32x4x2_t dequantized = vuzpq_f32(val_low, val_high);
-        vst2q_f32(reinterpret_cast<float *>(output.ptr()), dequantized);
+        // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255]
+        execute_window_loop(window_input_output, [&](const Coordinates & id)
+        {
+            // Get the input values
+            const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr() + id_batch[1] * _input->info()->strides_in_bytes()[3]);
+
+            const uint8x8_t  val_u8       = vld1_u8(input_ptr);
+            const uint16x8_t val_u16      = vmovl_u8(val_u8);
+            const uint32x4_t val_u32_low  = vmovl_u16(vget_low_u16(val_u16));
+            const uint32x4_t val_u32_high = vmovl_u16(vget_high_u16(val_u16));
+            float32x4_t      val_low      = vcvtq_f32_u32(val_u32_low);
+            float32x4_t      val_high     = vcvtq_f32_u32(val_u32_high);
+
+            // Dequantize -> (q / 255.0 * range) + min
+            val_low  = vmulq_f32(val_low, scaling);
+            val_high = vmulq_f32(val_high, scaling);
+            val_low  = vaddq_f32(val_low, vmin);
+            val_high = vaddq_f32(val_high, vmin);
+
+            const float32x4x2_t dequantized = vuzpq_f32(val_low, val_high);
+
+            // Store the dequantized values
+            auto output_ptr = reinterpret_cast<float *>(output.ptr() + id_batch[1] * _output->info()->strides_in_bytes()[3]);
+            vst2q_f32(output_ptr, dequantized);
+        },
+        input, output);
     },
-    input, output);
-}
+    min_max);
+}
\ No newline at end of file
diff --git a/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp b/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp
new file mode 100644
index 0000000..5e6c48f
--- /dev/null
+++ b/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp
@@ -0,0 +1,190 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h"
+
+#include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <algorithm>
+#include <arm_neon.h>
+#include <climits>
+#include <cstddef>
+
+namespace arm_compute
+{
+NEMinMaxLayerKernel::NEMinMaxLayerKernel()
+    : _input(nullptr), _output(nullptr), _mtx()
+{
+}
+
+void NEMinMaxLayerKernel::configure(const ITensor *input, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() < 3);
+    ARM_COMPUTE_ERROR_ON(output == nullptr);
+
+    TensorShape output_shape{ input->info()->tensor_shape() };
+    output_shape.set(Window::DimX, 2);
+    output_shape.remove_dimension(1);
+    output_shape.remove_dimension(1);
+
+    // Output auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
+
+    _input  = input;
+    _output = output;
+
+    // Configure kernel window
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output->info(), 0, 2);
+
+    update_window_and_padding(win, input_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEMinMaxLayerKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    const int x_start = window.x().start();
+    const int x_end   = window.x().end();
+
+    Window window_output;
+    window_output.use_tensor_dimensions(_output->info()->tensor_shape());
+    window_output.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    // Handle X dimension manually to split into two loops
+    // First one will use vector operations, second one processes the left over pixels
+    Window window_input(window);
+    window_input.set(Window::DimX, Window::Dimension(0, 1, 1));
+    window_input.collapse_if_possible(INEKernel::window(), 3);
+    window_input.set(3, Window::Dimension(0, 1, 1));
+
+    Iterator input(_input, window_input);
+    Iterator output(_output, window_output);
+
+    execute_window_loop(window_output, [&](const Coordinates & id_batch)
+    {
+        float32x2_t carry_min = vdup_n_f32(std::numeric_limits<float>::max());
+        float32x2_t carry_max = vdup_n_f32(std::numeric_limits<float>::lowest());
+
+        float carry_min_scalar = std::numeric_limits<float>::max();
+        float carry_max_scalar = std::numeric_limits<float>::lowest();
+
+        execute_window_loop(window_input, [&](const Coordinates & id)
+        {
+            int        x      = x_start;
+            const auto in_ptr = reinterpret_cast<const float *const>(input.ptr() + id_batch[1] * _input->info()->strides_in_bytes()[3]);
+
+            // Vector loop
+            for(; x <= x_end - 8; x += 8)
+            {
+                const float32x4x2_t pixels   = vld2q_f32(in_ptr + x);
+                const float32x4_t   tmp_min1 = vminq_f32(pixels.val[0], pixels.val[1]);
+                const float32x4_t   tmp_max1 = vmaxq_f32(pixels.val[0], pixels.val[1]);
+                const float32x2_t   tmp_min2 = vmin_f32(vget_high_f32(tmp_min1), vget_low_f32(tmp_min1));
+                const float32x2_t   tmp_max2 = vmax_f32(vget_high_f32(tmp_max1), vget_low_f32(tmp_max1));
+                carry_min                    = vmin_f32(tmp_min2, carry_min);
+                carry_max                    = vmax_f32(tmp_max2, carry_max);
+            }
+
+            // Process leftover pixels
+            for(; x < x_end; ++x)
+            {
+                const float pixel = in_ptr[x];
+                carry_min_scalar  = std::min(pixel, carry_min_scalar);
+                carry_max_scalar  = std::max(pixel, carry_max_scalar);
+            }
+        },
+        input);
+
+        // Reduce result
+        carry_min = vpmin_f32(carry_min, carry_min);
+        carry_max = vpmax_f32(carry_max, carry_max);
+        carry_min = vpmin_f32(carry_min, carry_min);
+        carry_max = vpmax_f32(carry_max, carry_max);
+
+        // Extract max/min values
+        const float min_i = std::min(vget_lane_f32(carry_min, 0), carry_min_scalar);
+        const float max_i = std::max(vget_lane_f32(carry_max, 0), carry_max_scalar);
+
+        auto out_ptr = reinterpret_cast<float *const>(output.ptr());
+
+        // Perform reduction of local min/max values
+        update_min_max(out_ptr, min_i, max_i);
+    },
+    output);
+}
+
+void NEMinMaxLayerKernel::reset()
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    float32x2_t reset_values = vdup_n_f32(0.0f);
+    reset_values             = vset_lane_f32(std::numeric_limits<float>::max(), reset_values, 0);
+    reset_values             = vset_lane_f32(std::numeric_limits<float>::min(), reset_values, 1);
+
+    Window window_output;
+    window_output.use_tensor_dimensions(_output->info()->tensor_shape());
+    window_output.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator output(_output, window_output);
+
+    execute_window_loop(window_output, [&](const Coordinates & id)
+    {
+        vst1_f32(reinterpret_cast<float *const>(output.ptr()), reset_values);
+    },
+    output);
+}
+
+void NEMinMaxLayerKernel::update_min_max(float *out_ptr, float min, float max)
+{
+    std::lock_guard<std::mutex> lock(_mtx);
+
+    const float32x2_t old_min = vld1_dup_f32(out_ptr);
+    const float32x2_t old_max = vld1_dup_f32(out_ptr + 1);
+    const float32x2_t new_min = vmin_f32(vdup_n_f32(min), old_min);
+    const float32x2_t new_max = vmax_f32(vdup_n_f32(max), old_max);
+
+    vst1_f32(out_ptr, vzip_f32(new_min, new_max).val[0]);
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp b/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp
index a596d83..bff79f0 100644
--- a/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp
@@ -23,9 +23,9 @@
  */
 #include "arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
@@ -35,14 +35,15 @@
 using namespace arm_compute;
 
 NEQuantizationLayerKernel::NEQuantizationLayerKernel()
-    : _input(nullptr), _output(nullptr), _min(nullptr), _max(nullptr)
+    : _input(nullptr), _output(nullptr), _min_max(nullptr)
 {
 }
 
-void NEQuantizationLayerKernel::configure(const ITensor *input, ITensor *output, const float *min, const float *max)
+void NEQuantizationLayerKernel::configure(const ITensor *input, ITensor *output, const ITensor *min_max)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() < 3);
 
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, DataType::U8, 0);
@@ -50,17 +51,20 @@
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
 
-    _input  = input;
-    _output = output;
-    _min    = min;
-    _max    = max;
+    _input   = input;
+    _output  = output;
+    _min_max = min_max;
 
     constexpr unsigned int num_elems_processed_per_iteration = 8;
 
     // Configure window
     Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-    update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), output_access);
+    AccessWindowStatic     min_max_access(min_max->info(), 0, 0, 2, min_max->info()->dimension(1));
+
+    // Update window and padding
+    update_window_and_padding(win, input_access, output_access, min_max_access);
     output_access.set_valid_region(win, input->info()->valid_region());
 
     INEKernel::configure(win);
@@ -72,36 +76,67 @@
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    Iterator input(_input, window);
-    Iterator output(_output, window);
+    Window window_input_output(window);
+    window_input_output.collapse_if_possible(INEKernel::window(), 3);
+    window_input_output.set(3, Window::Dimension(0, 1, 1));
 
-    const float32x4_t vmin             = vdupq_n_f32(*_min);
-    const float32x4_t inv_range        = vdupq_n_f32(1.0f / (*_max - *_min));
-    const float32x4_t quantization_max = vdupq_n_f32(255.0f);
-    const float32x4_t quantization_mul = vdupq_n_f32(256.0f);
+    Window window_min_max;
+    window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape());
+    window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1));
+    window_min_max.collapse_if_possible(INEKernel::window(), 1);
 
-    // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255]
-    execute_window_loop(window, [&](const Coordinates & id)
+    Iterator input(_input, window_input_output);
+    Iterator output(_output, window_input_output);
+    Iterator min_max(_min_max, window_min_max);
+
+    execute_window_loop(window_min_max, [&](const Coordinates & id_batch)
     {
-        float32x4x2_t val = vld2q_f32(reinterpret_cast<const float *>(input.ptr()));
-        // Map float values to range [0.0, 1.0]
-        val.val[0] = vsubq_f32(val.val[0], vmin);
-        val.val[1] = vsubq_f32(val.val[1], vmin);
-        val.val[0] = vmulq_f32(val.val[0], inv_range);
-        val.val[1] = vmulq_f32(val.val[1], inv_range);
+        // Get the min and max
+        float min = *(reinterpret_cast<const float *>(min_max.ptr()) + 0);
+        float max = *(reinterpret_cast<const float *>(min_max.ptr()) + 1);
 
-        // Quantize
-        val.val[0] = vmulq_f32(val.val[0], quantization_mul);
-        val.val[1] = vmulq_f32(val.val[1], quantization_mul);
-        val.val[0] = vminq_f32(val.val[0], quantization_max);
-        val.val[1] = vminq_f32(val.val[1], quantization_max);
+        // Saturate the result if min = max
+        if(min == max)
+        {
+            min = 0.0f;
+            max = 1.0f;
+        }
 
-        const uint32x4_t   val_u32_low  = vcvtq_u32_f32(val.val[0]);
-        const uint32x4_t   val_u32_high = vcvtq_u32_f32(val.val[1]);
-        const uint16x4x2_t val_u16      = vzip_u16(vmovn_u32(val_u32_low), vmovn_u32(val_u32_high));
+        const float32x4_t vmin             = vdupq_n_f32(min);
+        const float32x4_t inv_range        = vdupq_n_f32(1.0f / (max - min));
+        const float32x4_t quantization_max = vdupq_n_f32(255.0f);
+        const float32x4_t quantization_mul = vdupq_n_f32(256.0f);
 
-        const uint8x8_t quantized = vmovn_u16(vcombine_u16(val_u16.val[0], val_u16.val[1]));
-        vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), quantized);
+        // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255]
+        execute_window_loop(window_input_output, [&](const Coordinates & id)
+        {
+            // Get the input values
+            const auto    input_ptr = reinterpret_cast<const float *>(input.ptr() + id_batch[1] * _input->info()->strides_in_bytes()[3]);
+            float32x4x2_t val       = vld2q_f32(input_ptr);
+
+            // Map float values to range [0.0, 1.0]
+            val.val[0] = vsubq_f32(val.val[0], vmin);
+            val.val[1] = vsubq_f32(val.val[1], vmin);
+            val.val[0] = vmulq_f32(val.val[0], inv_range);
+            val.val[1] = vmulq_f32(val.val[1], inv_range);
+
+            // Quantize
+            val.val[0] = vmulq_f32(val.val[0], quantization_mul);
+            val.val[1] = vmulq_f32(val.val[1], quantization_mul);
+            val.val[0] = vminq_f32(val.val[0], quantization_max);
+            val.val[1] = vminq_f32(val.val[1], quantization_max);
+
+            const uint32x4_t   val_u32_low  = vcvtq_u32_f32(val.val[0]);
+            const uint32x4_t   val_u32_high = vcvtq_u32_f32(val.val[1]);
+            const uint16x4x2_t val_u16      = vzip_u16(vmovn_u32(val_u32_low), vmovn_u32(val_u32_high));
+
+            const uint8x8_t quantized = vmovn_u16(vcombine_u16(val_u16.val[0], val_u16.val[1]));
+
+            // Store the quantized values
+            auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr() + id_batch[1] * _output->info()->strides_in_bytes()[3]);
+            vst1_u8(output_ptr, quantized);
+        },
+        input, output);
     },
-    input, output);
+    min_max);
 }
diff --git a/src/runtime/NEON/functions/NEDequantizationLayer.cpp b/src/runtime/NEON/functions/NEDequantizationLayer.cpp
index f174367..a58b6e4 100644
--- a/src/runtime/NEON/functions/NEDequantizationLayer.cpp
+++ b/src/runtime/NEON/functions/NEDequantizationLayer.cpp
@@ -24,9 +24,7 @@
 
 #include "arm_compute/runtime/NEON/functions/NEDequantizationLayer.h"
 
-#include "arm_compute/core/Error.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/core/Validate.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
 using namespace arm_compute;
@@ -36,13 +34,13 @@
 {
 }
 
-void NEDequantizationLayer::configure(const ITensor *input, ITensor *output, const float *min, const float *max)
+void NEDequantizationLayer::configure(const ITensor *input, ITensor *output, const ITensor *min_max)
 {
-    // Configure kernels
-    _dequantize_kernel.configure(input, output, min, max);
+    // Configure kernel
+    _dequantize_kernel.configure(input, output, min_max);
 }
 
 void NEDequantizationLayer::run()
 {
     NEScheduler::get().schedule(&_dequantize_kernel, Window::DimY);
-}
+}
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEQuantizationLayer.cpp b/src/runtime/NEON/functions/NEQuantizationLayer.cpp
index 46b9d7d..a131c48 100644
--- a/src/runtime/NEON/functions/NEQuantizationLayer.cpp
+++ b/src/runtime/NEON/functions/NEQuantizationLayer.cpp
@@ -30,17 +30,20 @@
 using namespace arm_compute;
 
 NEQuantizationLayer::NEQuantizationLayer()
-    : _quantize_kernel(), _min_max_kernel(), _min(0.f), _max(0.f)
+    : _quantize_kernel(), _min_max_kernel(), _min_max()
 {
 }
 
 void NEQuantizationLayer::configure(const ITensor *input, ITensor *output)
 {
-    // Configure min-max kernel
-    _min_max_kernel.configure(input, &_min, &_max);
+    // Configure min-max kernel. _min_max tensor will be auto-configured within the kernel
+    _min_max_kernel.configure(input, &_min_max);
 
     // Configure quantize kernel
-    _quantize_kernel.configure(input, output, &_min, &_max);
+    _quantize_kernel.configure(input, output, &_min_max);
+
+    // Allocate min_max tensor
+    _min_max.allocator()->allocate();
 }
 
 void NEQuantizationLayer::run()
diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 4c449a7..806fc04 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -63,6 +63,36 @@
     }
 };
 
+/** Data set containing small 3D tensor shapes. */
+class Small3DShapes final : public ShapeDataset
+{
+public:
+    Small3DShapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 7U, 7U, 5U },
+                     TensorShape{ 27U, 13U, 37U },
+                     TensorShape{ 128U, 64U, 21U }
+    })
+    {
+    }
+};
+
+/** Data set containing small 4D tensor shapes. */
+class Small4DShapes final : public ShapeDataset
+{
+public:
+    Small4DShapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 7U, 7U, 5U, 3U },
+                     TensorShape{ 27U, 13U, 37U, 2U },
+                     TensorShape{ 128U, 64U, 21U, 3U }
+    })
+    {
+    }
+};
+
 /** Data set containing small tensor shapes. */
 class SmallShapes final : public ShapeDataset
 {
@@ -117,6 +147,36 @@
     }
 };
 
+/** Data set containing large 3D tensor shapes. */
+class Large3DShapes final : public ShapeDataset
+{
+public:
+    Large3DShapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 320U, 240U, 3U },
+                     TensorShape{ 383U, 653U, 2U },
+                     TensorShape{ 721U, 123U, 13U }
+    })
+    {
+    }
+};
+
+/** Data set containing large 4D tensor shapes. */
+class Large4DShapes final : public ShapeDataset
+{
+public:
+    Large4DShapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 320U, 123U, 3U, 3U },
+                     TensorShape{ 383U, 413U, 2U, 3U },
+                     TensorShape{ 517U, 123U, 13U, 2U }
+    })
+    {
+    }
+};
+
 /** Data set containing small tensor shapes for direct convolution. */
 class SmallDirectConvolutionShapes final : public ShapeDataset
 {
diff --git a/tests/validation/CPP/DequantizationLayer.cpp b/tests/validation/CPP/DequantizationLayer.cpp
index 1c7ec25..33096a1 100644
--- a/tests/validation/CPP/DequantizationLayer.cpp
+++ b/tests/validation/CPP/DequantizationLayer.cpp
@@ -32,23 +32,35 @@
 namespace reference
 {
 template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type>
-SimpleTensor<float> dequantization_layer(const SimpleTensor<T> &src, float min, float max)
+SimpleTensor<float> dequantization_layer(const SimpleTensor<T> &src, const SimpleTensor<float> &min_max)
 {
     // Create reference
     SimpleTensor<float> dst{ src.shape(), DataType::F32 };
 
-    const float range   = max - min;
-    const float scaling = range / 255.0f;
+    // Compute reference
+    const int width       = src.shape().x();
+    const int height      = src.shape().y();
+    const int depth       = src.shape().z();
+    const int stride_w    = width * height * depth;
+    const int num_batches = min_max.shape().total_size_upper(1);
 
-    for(int i = 0; i < src.num_elements(); ++i)
+    for(int k = 0; k < num_batches; ++k)
     {
-        dst[i] = (static_cast<float>(src[i]) * scaling) + min;
+        const float min     = min_max[k * 2 + 0];
+        const float max     = min_max[k * 2 + 1];
+        const float range   = max - min;
+        const float scaling = range / 255.0f;
+
+        for(int i = 0; i < stride_w; ++i)
+        {
+            dst[i + k * stride_w] = (static_cast<float>(src[i + k * stride_w]) * scaling) + min;
+        }
     }
 
     return dst;
 }
 
-template SimpleTensor<float> dequantization_layer(const SimpleTensor<uint8_t> &src, float min, float max);
+template SimpleTensor<float> dequantization_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<float> &min_max);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CPP/DequantizationLayer.h b/tests/validation/CPP/DequantizationLayer.h
index 3aae338..1a8adcf 100644
--- a/tests/validation/CPP/DequantizationLayer.h
+++ b/tests/validation/CPP/DequantizationLayer.h
@@ -36,7 +36,7 @@
 namespace reference
 {
 template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
-SimpleTensor<float> dequantization_layer(const SimpleTensor<T> &src, float min, float max);
+SimpleTensor<float> dequantization_layer(const SimpleTensor<T> &src, const SimpleTensor<float> &min_max);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CPP/QuantizationLayer.cpp b/tests/validation/CPP/QuantizationLayer.cpp
index d61e75a..0584d88 100644
--- a/tests/validation/CPP/QuantizationLayer.cpp
+++ b/tests/validation/CPP/QuantizationLayer.cpp
@@ -60,19 +60,48 @@
     // Create reference
     SimpleTensor<uint8_t> dst{ src.shape(), DataType::U8 };
 
-    // Compute min and max of the tensor using Min-Max layer
-    float min = 0.f;
-    float max = 0.f;
+    const int width       = src.shape().x();
+    const int height      = src.shape().y();
+    const int depth       = src.shape().z();
+    const int stride_w    = width * height * depth;
+    const int num_batches = src.shape().total_size_upper(3);
 
-    compute_min_max(src, &min, &max);
-
-    const float range = max - min;
-
-    for(int i = 0; i < src.num_elements(); ++i)
+    for(int k = 0; k < num_batches; ++k)
     {
-        // map values to range [0.0, 1.0]
-        const float normalized = (src[i] - min) / range;
-        dst[i]                 = static_cast<uint8_t>(std::min(255.0f, normalized * 256.0f));
+        // Compute min and max of the 3D tensor
+        float min = src[0];
+        float max = src[0];
+
+        // Look for min and max values
+        for(int i = 1; i < stride_w; ++i)
+        {
+            float val = src[i + k * stride_w];
+            if(val < min)
+            {
+                min = val;
+            }
+            if(val > max)
+            {
+                max = val;
+            }
+        }
+
+        // Saturate the result in case min = max
+        if(min == max)
+        {
+            min = 0.0f;
+            max = 1.0f;
+        }
+
+        const float range = max - min;
+
+        for(int i = 0; i < stride_w; ++i)
+        {
+            // map values to range [0.0, 1.0]
+            float       val        = src[i + k * stride_w];
+            const float normalized = (val - min) / range;
+            dst[i + k * stride_w]  = static_cast<uint8_t>(std::min(255.0f, normalized * 256.0f));
+        }
     }
 
     return dst;
diff --git a/tests/validation/NEON/DequantizationLayer.cpp b/tests/validation/NEON/DequantizationLayer.cpp
index 22d56ab..9bdba72 100644
--- a/tests/validation/NEON/DequantizationLayer.cpp
+++ b/tests/validation/NEON/DequantizationLayer.cpp
@@ -44,35 +44,56 @@
 {
 /** Tolerance for float operations */
 constexpr AbsoluteTolerance<float> tolerance_f32(0.001f);
+
+const auto DequantizationShapes = concat(concat(concat(datasets::Small3DShapes(),
+                                                       datasets::Large3DShapes()),
+                                                datasets::Small4DShapes()),
+                                         datasets::Large4DShapes());
+
 } // namespace
 
 TEST_SUITE(NEON)
 TEST_SUITE(DequantizationLayer)
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", DataType::U8)), shape, data_type)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(DequantizationShapes, framework::dataset::make("DataType", DataType::U8)), shape, data_type)
 {
+    TensorShape shape_min_max = shape;
+    shape_min_max.set(Window::DimX, 2);
+
+    // Remove Y and Z dimensions and keep the batches
+    shape_min_max.remove_dimension(1);
+    shape_min_max.remove_dimension(1);
+
     // Create tensors
-    Tensor src = create_tensor<Tensor>(shape, data_type);
-    Tensor dst = create_tensor<Tensor>(shape, DataType::F32);
+    Tensor src     = create_tensor<Tensor>(shape, data_type);
+    Tensor dst     = create_tensor<Tensor>(shape, DataType::F32);
+    Tensor min_max = create_tensor<Tensor>(shape_min_max, DataType::F32);
 
     ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
     ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(min_max.info()->is_resizable(), framework::LogLevel::ERRORS);
 
     // Create and configure function
-    float                 min = 0.f;
-    float                 max = 0.f;
     NEDequantizationLayer dequant_layer;
-    dequant_layer.configure(&src, &dst, &min, &max);
+    dequant_layer.configure(&src, &dst, &min_max);
 
     // Validate valid region
     const ValidRegion valid_region = shape_to_valid_region(shape);
     validate(src.info()->valid_region(), valid_region);
     validate(dst.info()->valid_region(), valid_region);
 
+    // Validate valid region of min_max tensor
+    const ValidRegion valid_region_min_max = shape_to_valid_region(shape_min_max);
+    validate(min_max.info()->valid_region(), valid_region_min_max);
+
     // Validate padding
     const PaddingSize padding = PaddingCalculator(shape.x(), 8).required_padding();
     validate(src.info()->padding(), padding);
     validate(dst.info()->padding(), padding);
+
+    // Validate padding of min_max tensor
+    const PaddingSize padding_min_max = PaddingCalculator(shape_min_max.x(), 2).required_padding();
+    validate(min_max.info()->padding(), padding_min_max);
 }
 
 template <typename T>
@@ -80,12 +101,14 @@
 
 TEST_SUITE(Integer)
 TEST_SUITE(U8)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType", DataType::U8)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(concat(datasets::Small3DShapes(), datasets::Small4DShapes()),
+                                                                                                                   framework::dataset::make("DataType", DataType::U8)))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), framework::dataset::make("DataType", DataType::U8)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(concat(datasets::Large3DShapes(), datasets::Large4DShapes()),
+                                                                                                                 framework::dataset::make("DataType", DataType::U8)))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_f32);
diff --git a/tests/validation/NEON/QuantizationLayer.cpp b/tests/validation/NEON/QuantizationLayer.cpp
index 5c2fab4..26657c4 100644
--- a/tests/validation/NEON/QuantizationLayer.cpp
+++ b/tests/validation/NEON/QuantizationLayer.cpp
@@ -44,12 +44,17 @@
 {
 /** Tolerance for quantization */
 constexpr AbsoluteTolerance<uint8_t> tolerance_u8(1);
+
+const auto QuantizationShapes = concat(concat(concat(datasets::Small3DShapes(),
+                                                     datasets::Large3DShapes()),
+                                              datasets::Small4DShapes()),
+                                       datasets::Large4DShapes());
 } // namespace
 
 TEST_SUITE(NEON)
 TEST_SUITE(QuantizationLayer)
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", DataType::F32)), shape, data_type)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(QuantizationShapes, framework::dataset::make("DataType", DataType::F32)), shape, data_type)
 {
     // Create tensors
     Tensor src = create_tensor<Tensor>(shape, data_type);
@@ -78,12 +83,14 @@
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEQuantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEQuantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(concat(datasets::Small3DShapes(), datasets::Small4DShapes()),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_u8);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEQuantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEQuantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(concat(datasets::Large3DShapes(), datasets::Large4DShapes()),
+                                                                                                             framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_u8);
diff --git a/tests/validation/fixtures/DequantizationLayerFixture.h b/tests/validation/fixtures/DequantizationLayerFixture.h
index 7543eb2..28d43cf 100644
--- a/tests/validation/fixtures/DequantizationLayerFixture.h
+++ b/tests/validation/fixtures/DequantizationLayerFixture.h
@@ -49,11 +49,8 @@
     template <typename...>
     void setup(TensorShape shape, DataType data_type)
     {
-        // Initialize random min and max values
-        rand_min_max(&_min, &_max);
-
-        _target    = compute_target(shape, data_type, _min, _max);
-        _reference = compute_reference(shape, data_type, _min, _max);
+        _target    = compute_target(shape, data_type);
+        _reference = compute_reference(shape, data_type);
     }
 
 protected:
@@ -63,28 +60,80 @@
         library->fill_tensor_uniform(tensor, 0);
     }
 
-    TensorType compute_target(const TensorShape &shape, DataType data_type, float min, float max)
+    template <typename U>
+    void fill_min_max(U &&tensor)
     {
+        std::mt19937                          gen(library->seed());
+        std::uniform_real_distribution<float> distribution(-1.0f, 1.0f);
+
+        Window window;
+
+        window.set(0, Window::Dimension(0, tensor.shape()[0], 2));
+
+        for(unsigned int d = 1; d < tensor.shape().num_dimensions(); ++d)
+        {
+            window.set(d, Window::Dimension(0, tensor.shape()[d], 1));
+        }
+
+        execute_window_loop(window, [&](const Coordinates & id)
+        {
+            const float n1 = distribution(gen);
+            const float n2 = distribution(gen);
+
+            float min = 0.0f;
+            float max = 0.0f;
+
+            if(n1 < n2)
+            {
+                min = n1;
+                max = n2;
+            }
+            else
+            {
+                min = n2;
+                max = n1;
+            }
+
+            auto out_ptr = reinterpret_cast<float *>(tensor(id));
+            out_ptr[0]   = min;
+            out_ptr[1]   = max;
+        });
+    }
+
+    TensorType compute_target(const TensorShape &shape, DataType data_type)
+    {
+        TensorShape shape_min_max = shape;
+        shape_min_max.set(Window::DimX, 2);
+
+        // Remove Y and Z dimensions and keep the batches
+        shape_min_max.remove_dimension(1);
+        shape_min_max.remove_dimension(1);
+
         // Create tensors
-        TensorType src = create_tensor<TensorType>(shape, data_type);
-        TensorType dst = create_tensor<TensorType>(shape, DataType::F32);
+        TensorType src     = create_tensor<TensorType>(shape, data_type);
+        TensorType dst     = create_tensor<TensorType>(shape, DataType::F32);
+        TensorType min_max = create_tensor<TensorType>(shape_min_max, DataType::F32);
 
         // Create and configure function
         FunctionType dequantization_layer;
-        dequantization_layer.configure(&src, &dst, &min, &max);
+        dequantization_layer.configure(&src, &dst, &min_max);
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(min_max.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Allocate tensors
         src.allocator()->allocate();
         dst.allocator()->allocate();
+        min_max.allocator()->allocate();
 
         ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!min_max.info()->is_resizable(), framework::LogLevel::ERRORS);
 
         // Fill tensors
         fill(AccessorType(src));
+        fill_min_max(AccessorType(min_max));
 
         // Compute function
         dequantization_layer.run();
@@ -92,43 +141,28 @@
         return dst;
     }
 
-    SimpleTensor<float> compute_reference(const TensorShape &shape, DataType data_type, float min, float max)
+    SimpleTensor<float> compute_reference(const TensorShape &shape, DataType data_type)
     {
+        TensorShape shape_min_max = shape;
+        shape_min_max.set(Window::DimX, 2);
+
+        // Remove Y and Z dimensions and keep the batches
+        shape_min_max.remove_dimension(1);
+        shape_min_max.remove_dimension(1);
+
         // Create reference
-        SimpleTensor<T> src{ shape, data_type };
+        SimpleTensor<T>     src{ shape, data_type };
+        SimpleTensor<float> min_max{ shape_min_max, data_type };
 
         // Fill reference
         fill(src);
+        fill_min_max(min_max);
 
-        return reference::dequantization_layer<T>(src, min, max);
-    }
-
-    /** Generate random constant values to be used as min and max for dequantization.
-     */
-    void rand_min_max(float *min, float *max)
-    {
-        std::mt19937                          gen(library->seed());
-        std::uniform_real_distribution<float> distribution(-10000.0, 10000.0);
-
-        const float n1 = distribution(gen);
-        const float n2 = distribution(gen);
-
-        if(n1 < n2)
-        {
-            *min = n1;
-            *max = n2;
-        }
-        else
-        {
-            *min = n2;
-            *max = n1;
-        }
+        return reference::dequantization_layer<T>(src, min_max);
     }
 
     TensorType          _target{};
     SimpleTensor<float> _reference{};
-    float               _min = 0.f;
-    float               _max = 0.f;
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>