FlexiGroBots - Fruit disease detection and counting
This service uses AI for precise fruit disease detection and fruit counting, enhancing agricultural productivity, optimizing crop management, and promoting good practices in modern agriculture.
This tool is a component of the European project, Flexigrobots.It consists of various models that together provide services for fruit disease detection and fruit counting. Users can utilize this service through Docker technology, accessible in a public GitHub repository associated with this application.
The fruit disease detection tool, improved since its initial presentation, uses three AI models across six steps to detect and classify fruit diseases. The workflow involves semantic segmentation using X-Decoder to identify relevant areas in UAV images, such as vineyards. It then focuses on the lower half of plants, where fruit is usually located, and creates patches for better cluster detection. Two options are available for fruit detection: a YOLOv8-seg model trained for specific fruits or an open-vocabulary Detic model for general fruit detection. The final step employs a YOLOv8-cls classifier to differentiate between healthy and unhealthy fruits, indicated by a color-coded segmentation mask.
Validation results show high accuracy in detection and classification, with the YOLOv8x-seg model achieving an F1 score of 0.82. Similarly, the YOLOv8x-cls model for classifying grape clusters shows an accuracy of 0.906. The tool’s functionality has been expanded to include fruit counting using image stitching and detection techniques, avoiding duplication in counts. This solution, while not real-time, provides results within a few hours, greatly aiding in accurate fruit quantity estimation before harvesting.