Research on a Lightweight Method for Maize Seed Quality Detection Based on Improved YOLOv8

Seeds are the most basic and important means of production for agriculture. During the production and processing of seeds, they may undergo potential mechanical damages and mildew alterations, which might jeopardize their germination viability. Hence, checking the quality of seeds before sowing is of paramount importance for the benefit of the sower and the safety of agricultural production. In order to achieve an efficient detection of maize seed quality, our experiment assembled a dataset composed of 2,128 seeds with four different health statuses of maize: healthy, broken, moth-eaten, and mildewed. In this thesis, we proposed a lightweight maize seed quality detection model for small objects based on improved YOLOv8: I-YOLOv8. Firstly, we introduced a multi-scale attention mechanism called EMA to efficiently retain information across channels and reduce computational load. Next, we chosen the SPD-Conv module for low-resolution images and small objects, and applied it to the backbone, which addressed the loss of fine-grained information and the less efficient learning of feature representations present in YOLOv8. Lastly, we reduced the large detection layer, which directed the network to pay more attention to the location, channel, and dimensional information of smaller objects, and we also replaced the loss function with WIoUv3. We validated our model using ablation studies and compared it with YOLOv5, YOLOv6, and YOLOv8. The mAP (Mean Average Precision) of the improved model I_YOLOv8 reaches 98.5%, which is 6.7% higher than YOLOv8. The average recognition time per image was 163.9fps, a boost of 5.2fps compared to YOLOv8. This study lays a theoretical foundation for the efficient, convenient, and rapid detection of maize quality, while also offering a technical basis for advancing automated maize quality detection means.