A Lightweight Multi‐Label Classification Model for Full‐Surface Defect Detection of Apples

ABSTRACT Fruit inevitably suffers damage during transportation. Detecting full‐surface defects on fruit before sale enhances commercialization capabilities. Currently, most machine vision‐based fruit defect classification systems only focus on detecting and recognizing part of the fruit surface. In this study, an auxiliary imaging device was designed to be used in conjunction with a camera to collect characteristic information about the entire surface of the apple. A full‐surface dataset of apples with a total of 1900 images was generated, containing seven types of feature labels (Normal, Disease, Insect‐bite, Rust, Crack, Scratch, and Bruise). A lightweight multi‐label classification network, MSFCLNet (multi‐scale focusing correlation lightweight network), was proposed to simultaneously detect all defect categories present on the apple surface. The multi‐scale focus module is utilized for visual feature extraction, while the tag association module is employed to mine and analyze the correlation between the labels of each category. Furthermore, label categories were balanced using asymmetric contrast loss to emphasize learning the important discriminative features of each label. Finally, multi‐label classification was achieved by employing associated classifiers. As a result, the acquisition of full‐surface image information, combined with a multi‐label classification approach, achieves complete and accurate detection and identification of fruit. MSFCLNet achieved a defect detection accuracy of 96.99%, with model parameters and FLOPs of 2.12 M and 827.50 M, respectively, striking a balance between accuracy and efficiency with superior performance. The method proposed in this study enhances the accuracy and completeness of fruit defect detection, thereby improving the overall quality of agricultural products.