Plant diseases significantly affect agricultural yields, leading to economic loss. The timely and accurate detection of potato leaf diseases is important to facilitate early intervention. Existing deep learning models struggle to detect minor variations in potato diseases due to uneven lighting condition, occlusions, complex background noise, and blur presence in the test images. To handle these problems, a multi-path, multi-attention, and lightweight transformer-based method incorporating shifted window multi-head self-attention (SW-MSA) with multi-scale channel attention and pixel attention is proposed in this paper. Additionally, an image super-resolution (ISR) module is also developed and incorporated with SW-MSA to further improve the disease detection accuracy by enhancing the quality of the original test images. Exhaustive experiments demonstrate that the proposed method performs superior compared to existing methods with an accuracy of 99.3% with lesser inference time. Confusion matrix and classification report-based evaluations affirm its better capability in identifying various types of diseases i.e., early blight and late blight with healthy classes. The proposed method provides a stable and accurate solution to real-time disease detection, facilitating early intervention and minimizing agricultural losses.
Full publication URL
