An improved YOLOv8 model for detecting damage types of maize kernels using proximal hyperspectral imaging

Rapid and accurate detection of maize kernel damage is crucial for breeding and food processing, but existing methods suffer from problems such as long detection times and weak discrimination capabilities. This study proposes a maize kernel damage type detection method that integrates hyperspectral imaging and deep learning algorithms. Addressing the challenge of effectively utilizing the large number of bands and information volume in hyperspectral imaging data, this study improves the YOLOv8 input layer to allow multi-channel hyperspectral input, introduces a Spatial-Channel Squeeze Attention mechanism (SCSA) to enhance the representation of key spectral-spatial features, and employs a Real-Time Detection Transformer (RT-DETR) as the detector to improve the localization and classification capabilities of complex damage, thereby achieving rapid and non-destructive identification of normal kernels and four types of damaged kernels. In this study, using 23 full bands as input, the improved YOLOv8 model achieved Precision, Recall, and mAP50 of 97.7%, 0.964, and 0.982, respectively, with an average detection time of 20.5 ms/image. Furthermore, to balance detection efficiency and accuracy, this study employed a sensitive band screening strategy, retaining only 5 sensitive bands to reduce the detection time to 10.3 ms/image while maintaining high detection accuracy (Precision 96.85%, Recall 0.953, mAP50 0.973). The improved YOLOv8 can detect maize kernel quickly, accurately and non-destructively, which is of great significance in crop detection. • A non-destructive maize kernel detection method using hyperspectral image is proposed. • Improved YOLOv8 target detection algorithm to adapt to hyperspectral image. • The improved YOLOv8 algorithm is used to detect five types of maize kernels. • The SCSA attention mechanism is used to distinguish different types of maize kernels.