Tomato Leaf Disease Detection using Enhanced Weight Assignment Based Loss Calculation Method

Plant diseases pose a significant threat to global food production and are a major challenge for crop management. Deep learning techniques have demonstrated remarkable success in plant disease detection, but most vision-based approaches rely on datasets captured in lab environments, limiting their applicability to real field images. The lack of large-scale non-lab datasets remains a significant challenge in enabling vision-based plant disease detection. This work focuses on leveraging sufficiently available in-lab datasets to enhance the effectiveness of plant disease recognition in real-world scenarios. This work proposes a region-reweighting-based approach that utilizes in-lab images for model training and in-field images for testing. The images are first partitioned into patches, and patch features are extracted using a pre-trained convolutional neural network (CNN). The clustering of these patches is performed to assign weights based on their similarity with other disease-containing patches. These assigned weights are considered during model training on the patches. Using this approach an accuracy of up to 92% is achieved on the disease class whose accuracy otherwise resulted in less than 90% on the state-of-the-art models. The work in this paper improves the identification accuracy of an individual disease class at the same maintaining overall accuracy as well.