Abstract: Aiming at the problems of poor accuracy of traditional apple leaf disease classification methods, an apple leaf classification algorithm CALDNet based on improved ConvNeXt is proposed. 3223 Network is designed to adjust the structure of the model, while jump connection and position coding are introduced to enhance the model’s ability to capture the space and to improve the stability of the training process, and Spatial Pyramid Pooling (SPP) is used to capture spatial features on different scales and enhance the model’s ability to adapt to large and small lesions; on the basis of ConvNeXtblock, G-ConvNeXtblock is designed to improve the depth convolution, and a Gabor filter is introduced as a convolution kernel to better capture texture information in the image. In order to improve the model’s ability to recognize a small range of apple leaf disease recognition ability, an enhanced channel and attention mechanism (enhanced CBAM) is designed. In the experiments, seven common leaf diseases (black-star disease, black-rot, brown-spot, mosaic disease, healthy, rust, gray-spot) are chosed as the main research subjects, and the experimental results by using the improved algorithm and other mainstream algorithms are compared. The experimental results show that the CALDNet model recognizes the leaf disease model with the precision rate, recall rate, and F₁ value of 97.58%, 97.54%, and 97.54%, respectively, compared with the original ConvNeXt model, which increased by 4.63,4.56 and 4.60 percentage points, solving the problems of poor precision of traditional apple leaf disease classification.

Key words: apple leaf disease, ConvNeXt, CNN, attention mechanism, deep learning