FENet: Frequency-Enhanced Network Based on AFFormer for Wood Surface Defect Detection

The bark is one of the major defects affecting the value of Eucalyptus veneer and must be accurately identified during detection. Currently, for bark defects exhibiting multiple shapes and colors, spatial-domain-based semantic segmentation models often encounter issues with semantic information loss in regions where the color or shape changes, which can lead to incomplete segmentation results. In contrast, utilizing texture features in the frequency domain can reduce the interference caused by variations in color and shape for the model, thereby enabling more effective identification of bark defects. Therefore, in this paper, a Frequency-Enhanced network based on Adaptive Frequency Transformer (AFFormer) is proposed. First, we propose an Inverted Depth-wise Separable Stem to extract more texture information by expanding the number of feature map channels through inverted depth-wise separable convolution. Moreover, we adopt the Rectangular Self-Calibration Module to refine the AFFormer as the backbone network. This enhances the ability to localize bark defects with different shapes and extracts frequency characteristics that are beneficial for semantic segmentation. Finally, the Frequency-Enhanced Channel Attention module enhances the frequency features for semantic segmentation and fuses the spatial-domain feature maps to recover the local details, thereby effectively improving the segmentation accuracy of multi-scale bark defects. Experimental results show that FENet outperforms existing semantic segmentation methods for segmenting bark defects.