MobileNet

Depthwise Separable Convolution is the core building block proposed by the MobileNet series; it splits standard convolution into two steps, depthwise and pointwise, and under 3×3 convolution it can usually reduce computation to one-eighth to one-ninth of the original.

A study systematically compared the performance of five CNN architectures—AlexNet, Inception, LeNet, VGG16 and MobileNet—in detecting twig blight of the Pequi tree (Caryocar brasiliense).

MobileNet is designed specifically for mobile and edge devices, greatly compressing the number of parameters through depthwise separable convolutions.

Depthwise separable convolution decomposes standard convolution into two steps, depthwise convolution and pointwise convolution, reducing computation and parameter count by about 8 to 9 times, at the cost of delaying inter-channel information interaction and slightly losing representational capacity.

Such AI models typically use convolutional neural networks (CNN) or lightweight vision models such as the MobileNet series, taking real-time frame sequences as input and outputting confidence heatmaps of suspected camera regions.

The power budget for TWS earphone chips is typically in the tens of milliwatts range, whereas inference for lightweight vision models often reaches hundreds of milliwatts, which could compress earphone battery life from several hours to under one hour.

Early edge AI relied mainly on lightweight convolutional neural networks (such as MobileNet and EfficientNet) for perception tasks, with parameter counts typically in the millions and inference times of only tens of milliseconds.

Depthwise separable convolution was popularized by Google in MobileNet and decomposes a standard convolution into two steps: a depthwise spatial convolution and a 1×1 pointwise channel-mixing operation.