基于改进RT-DETR的玉米缺陷粒检测算法OA

To address the challenges of easily losing small-scale defect features and insufficiently capturing high-frequency details in corn kernel defect detection,a lightweight and high-precision detection model(high-frequency defect capture detection Transformer,HDC-DETR)was proposed.Based on the real-time detection Transformer(RT-DETR),this model integrated the design concept of cross stage partial network(CSPNet)and high-frequency enhanced residual block(HFERB)in the backbone network to construct a high-frequency capture convolutional network(HCCNet)backbone.A diverse branch block(DBB)was introduced into the intermediate layers,based on which a diverse branch block-based C3 module(DBBC3)was constructed.During the upsampling stage of feature fusion,the content-aware reassembly of features(CARAFE)operator was employed to replace the traditional upsampling method.Performance comparisons between HDC-DETR and current mainstream detection models were conducted on a self-built corn kernel dataset,and ablation experiments were carried out to verify the effectiveness of each proposed module.The results demonstrate that on the self-built corn kernel test set,the mAP@50 and mAP@50-95 of HDC-DETR reach 96.48％and 68.51％,respectively,increasing by 2.25 percentage points and 2.23 percentage points compared to the original RT-DETR model.Furthermore,the parameter count and computational complexity of the model are reduced to 14.3 MB and 44.7 GFLOPs,decreasing by 28.14％and 21.58％compared to the baseline model,respectively.The proposed model effectively improves the detection accuracy of small-scale defects in corn kernels,providing reliable technical support for automatic agricultural product quality sorting systems,and holds significant practical application value in industrial quality inspection scenarios such as grain purchasing and preprocessing for deep processing.