Deep learning has been widely used in recent years for surface defect detection because of its excellent performance. However, current deep-learning-based approaches still remain a challenging problem in sawn lumber defect inspection because different lumber defects often keep similar texture and color since different growing environment, surface stains, etc. Meanwhile, the same defect often shows different characteristics. Furthermore, lumber defects have ambiguous boundaries or regions, and large-scale variations in size and shape. To address these problems, we have developed a deep Gaussian attention network via Deeplabv3+ for lumber surface defect segmentation. This network introduces an attention network via a transformer to capture the long-distance dependence for global information extraction, which can efficiently improve mis-segmentation since different lumber defects are very similar in some local regions. Furthermore, we introduce a Gaussian module into the channel attention module and positional attention module, respectively, to reassign and reactivate the minor semantic features for hard example mining so that all regions of defects can be activated except the key regions for efficient semantic segmentation. Finally, the activated global information and local information in astrous spatial pyramid pooling(ASPP) are integrated to achieve efficient feature extraction. Experimental results demonstrate the proposed network can efficiently address the ambiguous defect regions and irregular sizes and shapes of sawn lumber surface defect segmentation.

Lumber defect detection or segmentation is a crucial step in lumber industry. However, most existing semantic segmentation methods based on deep learning still remain challenging for lumber surface defect segmentation, because lumber defects often suffer from great variety, even if for the same defects. Moreover, lumber defects often have obscure boundaries and wide variations in size and shape. To address these issues, we develop a dual-branch discriminative feature cross-attention network (DBDFCNet) via U-Net for accurate defect segmentation. This network firstly introduces an improved Multi-scale Atrous Spatial Pyramid Pooling (MSASPP) to capture finer multi-scale features for enhancing the receptive field and addressing the wide range of defect sizes and shapes. Then, this network develops a dual-branch decoder which contains a binary branch and a semantic branch, in which the binary branch aims to determine defect localization and the semantic branch undertake the defect classification. In order to build the interaction between these two branches so that the binary branch to guide the semantic branch, we present a discriminative feature cross-attention module to increase inter-class distances and improving defect separability. Experimental results demonstrate that our method can effectively tackle the challenges of defect classification and edge segmentation for surface defect segmentation.