基于证据深度学习的苹果叶片病害分类网络OA

This paper addressed the issue that current apple leaf disease classification networks focused primarily on improving classification performance,while neglecting the reliability of predictions and the estimation of cognitive uncertainty.This limitation led to unreliable results and overconfidence when handling abnormal scenarios and compound disease samples.To tackle this problem,introduced evidential deep learning to explicitly model cognitive uncertainty and constructed a new classification network named EUCNet,using a Swin-Transformer backbone.Furthermore,to enhance the reliability of predictions,an evidence-based decision mechanism was proposed that leveraged uncertainty estimation to identify and flag unreliable samples for manual review.Experimental results demonstrated that EUCNet maintained classification performance comparable to mainstream networks under normal conditions while exhibiting significant advantages in various simulated abnormal scenarios.Specifically,its evidence-based decision mechanism effectively distinguished between high-confidence samples and those requiring rejection,with EUCNet outperforming mainstream classification networks on both types of samples.Furthermore,when confronted with compound disease samples,EUCNet achieved a successful detection rate of 86.57%,further validating the reliability and practical utility of the proposed method.