关键词: 深度学习, RT-DETR, 竹材加工, 缺陷检测, 特征融合, 模型轻量化

Abstract:

Objective: The existing bamboo strip surface defect detection methods on bamboo strips have problems such as large model parameters and poor performance in detecting small-scale defects, making it difficult to balance the dual requirements of high-precision and high real-time requirements in industrial scenarios. To address the problems, this paper proposes FCHM-DETR, a lightweight end-to-end detection model, for improving the comprehensive performance of bamboo strip surface defect detection. Method: Based on the RT-DETR network, a lightweight backbone Faster-CGLU was designed by integrating part convolutional and gated linear units (CGLU) to enhance the fusion capability of local and global features while reducing computational complexity. A spatial feature reconstruction pyramid network CGRFPN was constructed, which was combined with rectangular a self calibration module (RCM) and a dynamic interpolation fusion (DIF) mechanism to optimize the multi-scale feature fusion effect and improve the model's spatial perception ability for defects. A Haar Wavelet Downsampler was introduced to achieve efficient compression of feature maps through frequency-domain feature recombination, while fully retaining high-frequency detail information such as defect edges and texture mutations. In addition, we proposed an MPDIoU loss function, which improved the regression accuracy of defect bounding boxes and the convergence efficiency of the model by explicitly optimizing the distance constraints of bounding box diagonal corner points. Ablation experiments and comparative experiments with mainstream models were carried out on a dataset covering six common types of bamboo strip surface defects, including black knot, wormhole, mildew, crack, residual bamboo yellow and residual bamboo green, to verify the effectiveness of each module and the comprehensive performance of the proposed model. Result: The FCHM-DETR model achieved a mAP50 of 94.7%, which was 3.7% points higher than that of the baseline model RT-DETR-r18, with a significant performance improvement in detecting small-scale, low-contrast defects such as wormholes and cracks. Meanwhile, the parameter count and FLOPs of the model were reduced by 30.6% and 35.6% respectively compared with the baseline, and its inference speed reached 355 FPS, which can fully meet the frame rate requirements of real-time detection in industrial scenarios. Compared with mainstream defect detection models including YOLOv10m and Faster R-CNN, FCHM-DETR achieved a superior balance among three core dimensions of detection accuracy, lightweight level and inference speed, with outstanding comprehensive performance advantages. Conclusion: FCHM-DETR effectively breaks through the core technical bottlenecks of existing bamboo strip defect detection methods, and realizes the synergistic improvement of detection accuracy and industrial deployment adaptability. It can provide an end-to-end automated defect detection solution for the bamboo processing industry.

Key words: deep learning, RT-DETR, bamboo processing, defect detection, feature fusion, model lightweighting