Automatic identification of commodity label images using lightweight attention network

Recent research has raised interest in applying image classification techniques to automatically identify the commodity label images for the business automation of retail enterprises. These techniques can help enterprises improve their service efficiency and realize digital transformation. In this work, we developed a lightweight attention network with a small size and comparable precision, namely MS-DenseNet, to identify the commodity label images. MS-DenseNet is based on the recent well-known DenseNet architecture, where we replaced the regular planner convolution in dense blocks with depthwise separable convolution to compress the model size. Further, the SE modules were incorporated in the proposed network to highlight the useful feature channels while suppressing the useless feature channels, which made good use of interdependencies between channels and realized the maximum reuse of inter-channel relations. Besides, the two-stage progressive strategy was adopted in model training. The proposed procedure achieved significant performance gain with an average accuracy of 97.60% on the identification of commodity label images task. Also, it realized a 94.90% average accuracy on public datasets. The experimental findings present a substantial performance compared with existing methods and also demonstrate the effectiveness and extensibility of the proposed procedure. Our code is available at https://github.com/xtu502/Automatic-identification-of-commodity-label-images .