Classification of wild mushrooms based on improved ShuffleNetV2

This study introduced an improved CHE_ShuffleNetV2 model based on ShuffleNetV2 to address the classification challenge of wild mushrooms in a complex environment. The model incorporated a Cross Stage Partial (CSP) structure to simplify its complexity. Furthermore, it adopted Hybrid Dilated Convolution (HDC) to replace conventional convolution, enhancing the model’s recognition accuracy by expanding its receptive field. In addition, the ECA module was integrated to enhance the focus of the model on crucial feature information. The Hardswish activation function was employed instead of the ReLU activation function to reduce the number of parameters. The experimental results demonstrated that the enhanced model achieved improved accuracy, precision, recall, and F1-Score of 95.02%, 95.19%, 94.56%, and 94.00%, respectively, representing improvements of 2.81%, 3.82%, 3.08%, and 3.65%, correspondingly, over the original model. The enhanced model also reduced the parameters and FLOPs to 0.933 M and 104.08 M, respectively, representing reductions of 26.13% and 30.42% over the original model. Compared with commonly used lightweight models such as EfficientNet, DenseNet, and MobileNetV2, the CHE_ShuffleNetV2 model showed superior performance in solving the wild mushroom classification problem in complex environments, exhibiting its suitability for deployment on resource-constrained devices including mobile terminals.