Fusion of Transformer and Prototype Self supervised Identification of Apple Leaf Diseases

To tackle the persistent challenge of apple leaf diseases (ALD) identification, characterized by significant intra-class variation and subtle inter-class differences, this study presents an innovative model that integrates Transformer with Prototype Self-Supervised (FTPSS) learning. This model aims to significantly elevate the precision of ALD recognition, thereby enhancing disease management strategies in orchards. At the core of the FTPSS model lies the utilization of ResNet50 as the backbone network. This robust architecture is employed to extract multi-level feature maps from ALD images, capturing intricate details that are crucial for accurate disease identification. The extracted features are then processed through a novel encoder design, which integrates a Simplified Self-Attention (SSA) mechanism with Spatial Attention Guided Deformable Convolution (SAG-DC). This encoder, termed the Simplified Self-Attention and Deformable Convolution Transformer (SSADC-TF), facilitates the effective interaction and fusion of multi-level feature maps. By enhancing the model's sensitivity to irregular lesion areas within ALD images, SSADC-TF significantly boosts its ability to distinguish between different disease manifestations. To further refine the model's performance, a Prototype Self-Supervised (PSS) learning module is introduced. This module leverages two self-supervised loss functions: "Orthogonality" and "Clustering". The "Orthogonality" loss encourages the feature representations of different ALD classes to be orthogonal to each other, promoting a clear separation between classes and enhancing the model's discriminative ability. Meanwhile, the "Clustering" loss tightens the intra-class compactness, ensuring that variations within the same class do not undermine the model's robustness. Extensive experiments conducted on both standard and real-world image datasets demonstrate the remarkable effectiveness of the FTPSS model. On the standard image set, the FTPSS model achieves a recognition accuracy of 98.61%, marking a significant improvement of 5.15 percentage points over the baseline model. Similarly, on the real-world image set, the FTPSS model attains an accuracy of 98.73%, representing an enhancement of 4.49 percentage points compared to the baseline. These results underscore the FTPSS model's robust performance in identifying ALD, even in the presence of significant intra-class variation and subtle inter-class differences. The success of the FTPSS model can be attributed to its innovative integration of Transformer with Prototype Self-Supervised learning. By utilizing the powerful feature extraction capabilities of ResNet50 and the enhanced feature interaction and fusion provided by SSADC-TF, the model is able to capture complex details in ALD images, resulting in a 2.41 percentage point improvement. Furthermore, the introduction of PSS learning module helps to mitigate the semantic gap issue, ensuring that the model is able to generalize well to new, unseen ALD cases. The accuracy of ALD image recognition has increased by 1.8 percentage points.In conclusion, the FTPSS model presents a significant advancement in ALD recognition, with the potential to revolutionize disease management strategies in orchards. By automating the process of disease detection and providing precise, timely information, the FTPSS model can enable farmers to take swift action against ALD, thereby preserving the health and productivity of their orchards. This study contributes to the field of precision agriculture by demonstrating the power of advanced deep learning techniques in addressing complex agricultural challenges.