Leveraging YOLOv7 for Plant Disease Detection

Abstract

The agricultural sector contributes to 18.8% of India's Gross Domestic Product (GDP). With the increase in extreme climatic changes and constant deterioration of the quality of yield in the agricultural sector, detecting and treating plant diseases in their early stages is the need of the hour. At present, plant diseases are identified manually by examining them, which increases the time and decreases the efficiency and quality of the yield. This work focuses on providing a feasible solution to the problem of Plant Disease Detection. This work aims to develop a digital solution to this problem by training the fastest single-stage object detection model, YOLOv7, on the labeled PlantDoc Dataset. Since the PlantDoc dataset is small in size, data augmentation is performed. YOLOv7 achieves a significantly higher mean average precision of 71%. The size of the model is 75.1 MB, and the average time taken to detect an irregularity in an image is 6.8 ms. On account of the small size of the model and fast inference time, this model can be used for edge computing on devices such as satellites and drones to increase the yield produced.