Abstract:

In response to the problems of large parameter sizes, high computational complexity, long inference time, and difficulty in deployment on resource-limited mobile devices associated with existing tomato leaf recognition models, this paper proposes a lightweight recognition network called SG-ICA-MobileNetV3. Firstly, a sandglass structure was introduced to modify the inverted residual blocks of MobileNetV3Small, establishing feature transformations and skip connections in high-dimensional space to mitigate information loss and strengthen the model’s feature learning capability. Secondly, an improved coordinate attention mechanism was embedded, integrating global average pooling and max pooling to adaptively learn feature weights at different positions, enhancing the perception ability of diseased areas. Finally, the ReLU activation function was replaced with ELU to alleviate gradient vanishing and the issue of weight bias updated failure during model training, improving network convergence speed. The results showed that the model achieved a classification accuracy of 98.36% on the test set. It outperformed lightweight models such as MobileNetV3Small, MobileNeXt-1.0, MobileVit-S, and ConvNeXt-V2 in terms of parameter count, computational complexity, inference speed, and recognition accuracy, demonstrating strong generalization capabilities. This study result could provide algorithm support for fast and accurate identification of plant leaf diseases.

Key words: sandglass structure, SG-ICA-MobileNetV3, coordinate attention, tomato, leaf, disease recognition