MobileNet

MobileNet is a deep learning architecture developed by Google and specifically optimized to run on mobile and embedded devices. This architecture aims to significantly reduce computational cost and model size while preserving the performance of traditional convolutional neural networks (CNNs). MobileNet models are widely used in applications requiring low latency and low power consumption.

Depthwise Separable Convolution

The foundational building block of the MobileNet architecture is the depthwise separable convolution technique, which makes the standard convolution operation more efficient.

Computational Load of Standard Convolution

In classic convolutional layers, each filter is applied across the entire depth of the input. This requires a large number of multiply-add operations and increases computational cost.

Structure of Depthwise Separable Convolution

Depthwise separable convolution consists of two steps:

1. Depthwise Convolution: A separate convolution is applied to each input channel. This operation preserves the spatial dimensions of the input without changing the number of channels.
2. Pointwise Convolution (1×1): A 1×1 convolution is applied to transform the channel dimension. This step recombines the feature maps.

This structure can reduce computational load by approximately 8 to 9 times. The following image summarizes how depthwise separable convolution works.

Depthwise and Pointwise Convolution (Source =

Depthwise separable convolution requires significantly fewer parameters and computations compared to standard convolution.

MobileNet Variants

The MobileNet architecture has been developed over time through successive versions, each offering new improvements to the speed-performance trade-off.

MobileNetV1

• Based on the depthwise separable convolution architecture.
• Offers a scalable structure through hyperparameters: width multiplier (α) and resolution multiplier (ρ).
• Delivers high accuracy even in small models.

MobileNetV2

• Reduces information loss using structures called Inverted Residual Blocks.
• Enhances gradient flow through skip connections.
• Uses the ReLU6 activation function.

MobileNetV3

• The network architecture is automatically optimized using the Neural Architecture Search (NAS) method.
• Has two variants: Small and Large.
• Incorporates SE (Squeeze-and-Excitation) blocks.
• Represents the most advanced MobileNet version in terms of speed, accuracy, and efficiency.

Application Areas

MobileNet is a preferred architecture especially in resource-constrained environments. Some of its application areas include:

• Real-time image classification (mobile applications)
• Object detection and segmentation
• Face recognition systems
• Augmented reality (AR) applications
• Artificial intelligence on IoT devices