Improving Rice Leaf Disease Identification with Object Detection and Image Enhancement

Abstract

Highly accurate rice leaf disease identification using lighter models can ensure food security globally and also less crop loss by maximizing the rate of disease identifica tion in proper time. Our thesis focus is on the process where the quality of the image is enhanced and the diseased part is detected for predicting the disease in the output. Image enhancement refers to the process where noisy, low-contrast, and low-quality image is taken care of. The object detection mechanism detects which part of the im age contains disease. The identification process specifies the disease from the given input. The currently available disease detection model seems to have limitations, such as - high-quality training and testing images, requiring high computational power, images with biased backgrounds, etc. Our proposed technique works on these limita tions by including an image enhancement technique, Lite SR-GAN, that gives a better result even if we train our models on low-quality images. To make the computation less complex we have used lightweight architectures in the different phases (image enhancement, detection, disease identification) of the model. Next, to address the is sue of biased images, we have used an object detection technique, YOLO-s, to detect the diseased part of the image. In the final stage, we introduce a lightweight model, EfficientViT that, with the help of previous layers, works equally fine with pictures of all qualities and identifies the disease using less computation power with an accuracy of 97.615%, inference time 33.99ms per image, parameter count 4.56M, model size 6.95MB and flop count 33.56B.