A Pseudo-Label Refinement Method Based on Unified Local and Global Semantics for Semi-supervised Apple Leaf Disease Recognition

Semi-supervised learning can alleviate the reliance of fully supervised learning on expensive large-scale labeled data, which is particularly important in apple leaf disease recognition where expert experience is critical. However, due to the characteristics of apple leaf diseases, which have small interclass differences and large intraclass differences, leading to more noise in the pseudo labels, thereby preventing the generalized semi-supervised methods from being directly applied to apple leaf disease recognition. So, a novel semi-supervised recognition method which called ULG-Match is proposed in this paper, which fully utilizes both local and global information to enhance the performance of semi-supervised apple leaf disease recognition. Firstly, a region-based local feature extraction method is designed to extract more reliable local feature representations; secondly, a pseudo-label generation based on local perception enhancement module is proposed to generate more accurate pseudo-labels by unifying local and global semantic information. Finally, the proposed ULGMatch was compared with several SOTA methods, achieving improvements of 1.82% and 2.31% on the internal dataset apples and the publicly available dataset plantvillage, respectively.