Surface defect detection method for shawo radishes based on improved RT-DETR

Surface defect detection of shawo radishes is challenged by complex defect characteristics, large variations in defect scale, and insufficient feature representation, which can limit automated quality inspection and grading performance.

To address these challenges, this study proposes an improved RT-DETR-r18 framework based on a frequency-spatial collaborative modeling strategy. The proposed model incorporates a Wavelet Transform Block (WT_Block), an Attention-based Intra-scale Feature Interaction with High-Low Attention module (AIFI-HiLo), and a VoVGSCSP-PConv Cross-Scale Feature Fusion Module (VP_CCFM) to enhance feature extraction, feature interaction, and multi-scale representation.

Experiments conducted on a self-constructed shawo radish surface defect dataset showed that the proposed method achieved a Precision of 94.5%, a Recall of 81.2%, and a mean Average Precision at an Intersection over Union threshold of 0.5 (mAP@50) of 85.3%, representing improvements of 2.9, 2.8, and 1.9 percentage points over the baseline RT-DETR-r18 model, respectively. Meanwhile, the parameter count and computational complexity were reduced by 7 and 12.7%, respectively. Comparative experiments further showed that the proposed model outperformed several mainstream object detection methods in overall detection performance.

These results indicate that the proposed frequency-spatial collaborative framework can improve the accuracy, robustness, and efficiency of shawo radish surface defect detection, providing technical support for automated quality inspection and intelligent grading of root vegetable products.