MFWSD-YOLO: Lightweight Multi-scale Feature-fusion Wood Surface Defect Detection Algorithm

Abstract

Wood surface defect detection confronts critical challenges including cross-scale feature extraction, excessive parametric burden, and inadequate small-target recognition. This study proposes MFWSD-YOLO, a lightweight multi-scale feature fusion detection algorithm to address these limitations. The algorithm introduces an adaptive downsampling module utilizing dual-path parallel processing to preserve spatial information, designs a shared convolution detection head enabling efficient cross-scale feature interactions, proposes a progressive feature integration block strengthening multi-scale semantic fusion, and embeds a local attention mechanism enhancing spatial modeling precision. Experimental validation demonstrates substantial enhancements, achieving mAP@0.5 and mAP@0.5:0.95 improvements of 8.90% and 5.17% respectively over baseline YOLOv12n. Concurrently, efficiency gains include 52.73% parameter reduction, 33.33% computational complexity decrease, and 50.94% model size compression, maintaining 195.6 frames per second inference capability. Cross-dataset validation substantiates robust generalization across diverse wood defect scenarios and industrial applications. These advances establish an effective computational solution for automated wood quality inspection within intelligent manufacturing environments.