Improved YOLO-X Model for Tomato Disease Severity Detection using Field Dataset

Abstract

In the past decade, the field of automatic plant disease detection has undergone significant complexity. Advancements in convolutional neural network in deep learning have enabled the rapid and precise detection of ailments, facilitated the development of effective treatments and ultimately led to higher crop yields. One of the most challenging scenarios in plant disease occurs when multiple diseases manifest on a single leaf, exacerbating the difficulty of diagnosis due to overlapping symptoms. This study addresses these challenges by employing an enhanced YOLO-X model for detection tomato leaf diseases. The technique presented here enhances the Spatial Pyramid Pooling layer in order to extract valuable features from training data of various sizes more efficiently. We were able to increase the model's ability to identify a broader spectrum of disease symptoms by concatenating variables from multiple layers and varying sizes. In addition, we incorporate a large number of connections to increase the generalizability of the design. The application of an IoU-based (Intersection over Union) regression loss function increases the convergence of the network and the precision of the detection. For experimentation, we created a customized dataset consisting of 1220 tomato plant leaf images from various farms in Southern part of India, encompassing overlapping diseases and varying degrees of severity. The dataset includes images of healthy leaves as well as different severity levels of tomato leaf curl and tomato leaf mold stress on a single leaf. Our suggested improved SPP-based YOLO-X model beats the original YOLO-X model, according to experimental findings, which show an improvement in test dataset accuracy and a 73.42% mean Average Precision on field-collected dataset.