Lightweight Network Architecture Optimization of Deep Learning Models

With the widespread application of deep learning in image classification tasks, the demand for lightweight convolutional neural networks (CNNs) has been steadily increasing in resource-constrained environments such as mobile devices and embedded systems. To balance classification accuracy and computational efficiency, numerous lightweight network architecturessuch as MobileNet, ShuffleNet, SqueezeNet, and EfficientNethave introduced various techniques, including depthwise separable convolution, inverted residual blocks, channel shuffle, Fire modules, and compound scaling. This paper proposes an innovative lightweight network architecture, LwNet, which integrates and customizes modules based on these classical structures to enable more efficient feature extraction and information transmission. Specifically, we design a module named LightBlock, which combines 11 convolution expansion, depthwise separable convolution, grouped convolution, and channel shuffle operations. The module also incorporates residual connections and attention mechanisms to further enhance model performance. Experimental results demonstrate that LwNet achieves excellent classification accuracy on the CIFAR-10 dataset, offering a superior balance between accuracy and efficiency compared to other mainstream lightweight networks. This study provides a novel perspective for designing lightweight neural networks and offers reproducible code and experimental data to support further research in related fields.