Abstract: Aiming at  the problems of high computational complexity, large parameter size, and limited deployability of existing models in the  image recognition tasks of pepper diseases and pests,  we  proposed a lightweight  disease and pest recognition method based on an improved ResNet34. Firstly,  depthwise separable convolution structures were introduced to replace conventional standard convolutions, which significantly reduced model parameter count and computational cost while maintaining feature extraction capability. Secondly, a convolutional block attention mechanism was integrated to enhance the model’s ability to focus on critical region features of lesions from both channel and spatial dimensions, and  improve the discriminative power of feature representations. In order to verify the effectiveness of the proposed method, systematic experiments were conducted on a pepper  disease  and pest dataset as well as the public PlantVillage tomato disease dataset. Experimental results show that the improved model achieves significant improvements in  multiple evaluation metrics, with a recognition accuracy of 98.36%, which is  5.64% higher than  the original ResNet34. In comparative experiments with models such as MobileNetV2, GoogLeNet, AlexNet, VGG16, and ResNet18, the proposed method performs better  in accuracy, precision, recall, and F₁ values. Furthermore, the parameter count of proposed model is  only 2.73 M, and the  model volume is  10.62 MiB, which has good  lightweight characteristics and deployment applicability  while maintaining high recognition accuracy.

Key words: pepper disease and pest, ResNet34 network, depthwise separable convolution, attention mechanism, deep learning