Improving Network Structure for Efficient Classification Network Based on MobileNetV3

Abstract

MobileNetV3 employs depthwise separable convolutions to construct lightweight deep neural networks, but its use of neural architecture search and the Squeeze-and-Excitation (SE) attention mechanism introduces considerable computational overhead. To enhance classification performance while reducing computational cost, this paper proposes an improved model, ISNECA-MobileNetV3. The model enhances the bottleneck structure of MobileNetV3 to improve accuracy and embeds an Efficient Channel Attention (ECA) module to strengthen feature extraction with minimal overhead. Furthermore, TrivialAugment data augmentation and the SiLU activation function are adopted to boost overall performance. Experimental results on a waste classification dataset show that ISNECA-MobileNetV3 achieves a 4.1% improvement in accuracy and reduces model parameters by 2.4M compared to the original MobileNetV3, demonstrating significant gains in both inference speed and classification accuracy.