Wood Surface Discoloration Defect Detection Based on Improved HRNet

Discoloration defect detection is a significant task in the wood industry for ensuring the appearance and quality of wooden products. However, discoloration defect detection still faces some challenges due to the low contrast, blurred boundary, and slender shape of discoloration defects. To overcome these challenges, we propose a multi-branch feature encoding HRNet(MFE-HRNet) for wood surface discoloration defect detection. The network maintains high-resolution features and builds a parallel multi-resolution architecture to prevent the dilution of low-contrast discoloration defect features. Through repeated cross-fusion of multi-scale features, the network enhances the global representation of blurry features. In order to effectively preserve the detail features of the defect edge, a multi-branch feature encoding module (MFEM) is proposed as the encoding layer of the network. In the MFEM, multiple branches concurrently extract features, generating feature maps with complementary characteristics, thereby enhancing the ability of network to recognize fine details of defects. Experimental results verify the effectiveness of our method in detecting discoloration defects on wood surface.