Lightweight Model for Apple Leaf Disease Recognition Based on Haar Wavelet Downsampling and Focal Loss Function

Aiming at the problems of large number of parameters and high computational cost in disease recognition models, and the decline in model accuracy caused by image noise interference and sample class imbalance in apple leaf disease recognition, this paper proposes an improved model combining Haar wavelet down-sampling and focal loss function. The improved model uses MobileNetV3-Small as the base model and enhances its resistance to image noise in apple leaf disease recognition by applying Haar wavelet downsampling instead of the traditional downsampling method. Additionally, the focal loss function replaces the cross-entropy loss function, improving the model’s accuracy in recognizing minority classes of diseases. Experimental results show that the improved model, MobileNet-HWD-FL, achieves a recognition accuracy of 96.42% in the apple leaf disease recognition task, with 1.52M parameters and 68.19M FLOPs. Compared with classical models such as MobileNetV2, EfficientNet-B0, ShuffleNet_v2_1.5×, and DenseNet121, the improved model offers higher accuracy and fewer parameters. The improved method proposed in this paper provides an effective approach for lightweight identification of apple leaf diseases and other plant diseases.