A Lightweight Residual Dilated CNN–Transformer Framework for Efficient Rice Leaf Disease Classification

Although rice is a staple food for more than half of the world's population, several illnesses represent a severe threat to rice farming, lowering yields by as much as 70%. Human visual examination, the basis of traditional disease diagnosis, has several limitations, including subjectivity, inconsistent results, and the inability to cover large agricultural areas. So far, using deep learning to automate plant disease diagnosis has been effective, but there are still numerous problems to address. Some of these include that they are difficult to understand, require a lot of computer power, and are not well-suited for field use. This study proposes an innovative hybrid deep learning framework called RiceLeafCNN-Transformer, which integrates transformer architectures with convolutional neural networks to address existing limitations. A proposed method for representing features at multiple scales enhances computing speed by combining dilated convolutions, depth wise separable convolutions, and squeeze-excitation blocks. We can accomplish global contextual modelling and hierarchical feature extraction by combining a transformer encoder with a lightweight convolutional neural network (CNN) backbone. Proposed model shows state-of-the-art performance in its rigorous experimental validation using three datasets of rice leaf disease. Dataset A has 18,445 photos, and Dataset B has 11,790 photos. It does this by keeping inference times between 8.3 and 9.9 ms and getting 99.2% accuracy. The proposed strategy can help farmers and agricultural specialists save time and effort by linking theoretical lab performance to real-world use. This way, they get information that is both useful and relevant.