Recognition of pear leaf disease under complex background based on DBPNet and modified mobilenetV2

Abstract

AbstractGiven the challenge of pear leaf disease recognition caused by uneven illumination, overlapping leaves, and other green plants in the background, a two‐stage strategy‐based framework for pear leaf segmentation and disease classification is proposed. Initially, a double branch polymerization net fusing features of the low‐level feature branch and semantic branch is constructed to extract the target diseased pear leaf and eliminate background interference. Then an improved lightweight neural network (Inverted‐Inception Efficient‐Excitation‐and‐Filtering‐Bottleneck MobileNet‐v2, I2EMv2) is used to capture multi‐scale lesion information and ainhibit invalid feature channels and filter redundant features for the final classification. The experimental results show that the proposed framework can accurately extract the pear leaf region with complete boundary from the complex background, maximize the retention of lesion information, and achieve high‐precision pear leaf disease identification. The mean absolute error, F‐measure, and intersection over union of leaf segmentation are 0.027, 0.947, and 0.880, respectively, and the average recognition accuracy of leaf disease is 92.05%. Compared with others, the method proposed in this paper has superior performance on segmentation and classification, which provides a reference for pear leaf disease classification under complex background.