Optimized wheat seed classification using YOLO with morphological image feature enhancement

Abstract The article presents a superior computer vision system to detect and grade wheat seeds. It is concerned about the integration of deep-learning detection and conventional image processing methods to enhance the overall classification accuracy. The current methods of seed classification in wheat seeds have been found to lack proper visibility of the features particularly in low contrast images, small defects, or overlapping of seeds, and also, irregular lighting situations. The conventional feature extractors are weak, and the deep models by itself fail in cases where the morphological features like grooves, cracks, and shriveling are not well pronounced. To overcome these constraints, the proposed YOLO-Integrated Morphological Feature Enhancement Pipeline (Y-MFEP) uses dilation, erosion, opening, closing and top-hat transformations to enhance structural features and detect them using YOLO. The fused images are the improved feature maps and the original images, which allows the YOLO to identify finer seed variations more accurately. Such a hybrid pipeline enhances the visibility of the edges, defects, and texture uniformity without sacrificing the real-time detection performance. The given method is used to grade the quality of wheat at agricultural processing and procurement centers automatically. It guarantees quick, stable, and high-scaling classification of fit, broken, shriveled, and infected seeds. The results demonstrate that Y-MFEP has a major advantage of improving the accuracy, mAP and defect-detection sensitivity, which creates a more dependable and automated wheat quality measurement. The classification accuracy (85–95%), defect sensitivity index (0.775), edge clarity score (78–85), intersection over union (75–82%), and small object detection rate (75–90%) is reached in the proposed method.