Resumen
Cassava is a typical staple food in the tropics, and cassava leaf disease can cause massive yield reductions in cassava, resulting in substantial economic losses and a lack of staple foods. However, the existing convolutional neural network (CNN) for cassava leaf disease classification is easily affected by environmental background noise, which makes the CNN unable to extract robust features of cassava leaf disease. To solve the above problems, this paper introduces a transformer structure into the cassava leaf disease classification task for the first time and proposes a transformer-embedded ResNet (T-RNet) model, which enhances the focus on the target region by modeling global information and suppressing the interference of background noise. In addition, a novel loss function called focal angular margin penalty softmax loss (FAMP-Softmax) is proposed, which can guide the model to learn strict classification boundaries while fighting the unbalanced nature of the cassava leaf disease dataset. Compared to the Xception, VGG16 Inception-v3, ResNet-50, and DenseNet121 models, the proposed method achieves performance improvements of 3.05%, 2.62%, 3.13%, 2.12%, and 2.62% in recognition accuracy, respectively. Meanwhile, the extracted feature maps are visualized and analyzed by gradient-weighted class activation map (Grad_CAM) and 2D T-SNE, which provides interpretability for the final classification results. Extensive experimental results demonstrate that the method proposed in this paper can extract robust features from complex non-balanced disease datasets and effectively carry out the classification of cassava leaf disease.