Accurate Identification of Leaf Disease using Yolov4 Algorithm

Abstract

Aim: This study aims at comparing the effectiveness of the YOLOv4 algorithm against models based on Convolutional Neural Networks in order to increase the accuracy in disease identification. Materials and methods: For this investigation, two groups were tested. It involved training and evaluating the YOLOv4-based leaf disease detection model deployed in Group 1 on 26 samples of diseased and healthy leaf images. This was done while incorporating a number of environmental factors including variable illumination, occlusions, and orientations. Pre-training, contrast enhancing and sharpening of edges were performed on the images to enhance the accuracy of detection. The same 26 images from the given set are taken for training and testing of the CNN-based Group 2's leaf disease classifier. It had extracted features automatically with the help of convolutional layers. Accuracy, precision, recall, F1 score were compared to evaluate the classification of YOLOv4, CNN. A threshold was set up to observe the result, which was 0.05% and also 95% confidence interval with G Power of 80%. Result: The YOLOv4 model outperforms from the CNN model, where the CNN achieved an accuracy of 87.00%, with a processing time of 2.30 seconds (p < 0.05), while the YOLOv4 model achieved an accuracy of 96.00% with a little inference time, that is 0.65 seconds. The statistical analysis also proved the performance and effectiveness of YOLOv4 in real-time leaf disease detection. Conclusion: The outcome of the paper is that the advanced delicacy and fast conclusion time of YOLOv4 make it a better option than CNN for real-time splint complaint discovery. Using YOLOv4 in perfect husbandry will reduce crop losses, complaints will be identified at an early stage, and total agricultural productivity will increase.