mobilnet

MobileNet is a family of lightweight convolutional neural networks designed for efficient on-device vision tasks on mobile and embedded devices. The original MobileNetV1, introduced in 2017 by Google researchers, reduces computation and memory needs relative to traditional CNNs. It achieves this mainly with depthwise separable convolutions, which factor a standard convolution into a depthwise spatial filter for each input channel and a 1x1 pointwise filter that combines channels. The architecture also uses width and resolution multipliers, known as alpha and rho, to trade accuracy for speed and model size.

MobileNetV2, released in 2018, improves the design with inverted residuals and linear bottlenecks. Each block expands channels with a lightweight 1x1 convolution, applies depthwise convolution, and projects back to a smaller set of channels. This structure maintains representational power while keeping computation low, often delivering better accuracy at similar latency than MobileNetV1.

MobileNetV3, introduced in 2019, combines neural architecture search with human-designed blocks to optimize performance for mobile CPUs. It yields two main models, Large and Small, and uses lightweight attention modules and efficient nonlinearities to improve latency and accuracy on common mobile hardware.

Applications and impact: MobileNet architectures are widely used as compact backbones for image classification, object detection, segmentation, and on-device vision tasks. They appear in major frameworks and tooling, such as TensorFlow Lite, and remain common baselines for mobile-vision research and deployment.