Improved Lightweight Wood Defect Detection Algorithm Based on Yolov8

Aiming at the problems of large parameters and complex calculation of wood defect detection model, a lightweight model CCKD Ours was proposed. The CSPPC module integrates multi-scale cavity convolution and pooling, expands the receptive field and compresses the feature dimension, reducing the computational burden of defects such as cracks and dry joints. CCFM module adopts bidirectional channel attention suppression redundancy feature, and uses 1 × 1 convolutional residual fusion instead of splicing to reduce the amount of calculation. Combined with knowledge distillation, transfer YOLOv8m teacher model knowledge based on C2f_CBAM, ADown sampling and BiFPN structure to improve the performance of student models. Experiments show that the CCKD Ours parameter quantity decreases by 52% to 1.4 M, and the calculation quantity decreases by 38.3% to 5.0GFLOPs mAP@0.5 Increase by 1.6%, 2.0% and 1.5% respectively to achieve the balance between lightweight and precision.