Surface Defect Detection of Wooden Boards Based on Multi-Modal Fusion

Automated wood defect detection is of great significance for improving wood utilization. However, the stains on the wood surface resemble defect features and it is difficult to distinguish the appearance of wane from defect-free regions. Relying solely on machine vision for extracting surface features of wood makes it difficult to achieve accurate defect detection. To address these challenges, we propose MMDNet, a multimodal detection network that combines point cloud and image data. By adaptively fusing point cloud and image features at multiple stages, the network effectively highlights wood surface defect regions. Additionally, we introduce an Atrous Spatial Pyramid Pooling (ASPP) module into the network, which expands the network's receptive field and enables a comprehensive perception of the defects. Furthermore, we employ a deep supervision strategy to encourage the network to learn discriminative feature representations, enhancing the model's ability to differentiate defect regions. Experimental results validate the effectiveness of our method in detecting defects that are similar to the background and reducing interference from stains.