Rectifying the extremely weakened signals for cassava leaf disease detection

The performance of neural networks is heavily dependent on the integrity of the feature signals. As these signals are extracted and transmitted, they tend to weaken, which can negatively affect their ability to represent and utilize semantic information, particularly in weakly supervised learning tasks. This study aims to address hidden and severely weakened signals. To address the underlying causes, the rectification block of the third stage of PR-ArsenicNetPlus (Positive Rectified ArsenicNetPlus Neural Network) was modified to include a nonlinear fitting method based on the variant Hölder inequality. This method adjusts the magnitude and angular frequency of an extremely weak signal, and its effectiveness is evaluated using Parseval’s relationship. When tested on cassava leaf disease datasets, the proposed method significantly improved the prediction accuracy in 7-fold cross-validation, achieving an accuracy of 96.18 %, a loss of 1.373, and an F1-score of 0.9618. These results outperformed those of ResNet-101, EfficientNet-B5, RepVGG-B3g4, and AlexNet.