Non-exemplar class-incremental learning for continual plant diagnosis

Deep learning has been widely applied as a general technique for image classification in plant diagnosis. Despite the impressive performance verified by individual classification tasks, deep learning networks suffer from forgetting the knowledge of old-type when updating the input stream by new disease samples in the continual plant diagnosis. Recently, rehearsal-based class-incremental learning approaches for plant disease classification have been proposed to mitigate the effects of old-type forgetting. These methods stored parts of leaf images of old disease, then replayed old exemplars and trained jointly with the new disease data in a class-incremental task. However, privacy issues and a considerable amount of memory limit the application of these rehearsal-based methods. In this paper, we investigate non-exemplar class-incremental learning schemes for plant diagnosis to address the catastrophic forgetting problem without requiring extra memory space for stored exemplars. We introduce a new non-exemplar class-incremental learning scheme, NeCILPD, for continual plant diagnosis. In particular, we propose an improved self-supervision learning algorithm and a novel prototype inversion constraint strategy to mitigate the problem of prototype shifts, in order to further improve the performance of few-shot class-incremental learning tasks. Experimental results confirmed the effectiveness of the proposed class-incremental learning approach. Specifically, the proposed class-incremental learning scheme achieved 70.27% accuracy and 17.80% forgetting measure in the incremental classification of 30 categories, outperforming the current SOTA method, which attained 63.80% accuracy and a forgetting measure of 24.80%. The impressive performance of the proposed non-exemplar class-incremental learning scheme provides a reliable tool for continual plant diagnosis, laying a solid foundation for agricultural applications.