Based on the Depth-Separable Convolution of Lightweight Networks

Currently, the ubiquitous 1×1 convolution operation makes the network face the challenge of high computational complexity. At the same time, many models are dependent on data, and there are obvious differences between small sample data, which reduces the classification ability of the model. To this end, this paper proposes a lightweight network based on depth-wise separable convolution, which improves the classification ability of the model by reducing the complexity of the network, reduces the model dependence, and improves the classification accuracy by improving the attention and loss function. Firstly, a significant reduction in computational complexity is achieved by replacing the traditional 1×1 convolution with a grouped convolution. A cross-channel feature weighting module is introduced to overcome the limitation of grouped convolution in terms of information exchange between channels. In addition, to effectively deal with the data sample class imbalance problem, this research introduces an attention mechanism that leverages channel and feature space equalization to effectively capture global information. Additionally, it formulates a loss function that incorporates both intra- and inter-class metrics, aiming to enhance model accuracy. Finally, through validation on multiple datasets, the methods in this study show their efficiency and superiority.