WFF-Net: Trainable weight feature fusion convolutional neural networks for surface defect detection

Deep learning-based surface defect segmentation (SDS) technique is widely used in the field of surface defect detection (SDD) for its high accuracy and robustness. However, the deep learning-based surface defect segmentation method suffers from the interference from the semantic difference between the high-dimensional features and the low-dimensional features during the process of multi-scale feature fusion, which will bring additional noise to the network and thus affect the detection accuracy. For this drawback, this paper investigates a new weight-based feature fusion method, which aims to reduce the semantic differences and information redundancy after multiscale feature fusion in the process of coupling high-dimensional semantic features with low-dimensional semantic features. First, the WFF feature fusion method is proposed in the multi-scale feature fusion stage, which uses a learnable Gate module to assign weight coefficients to neighboring features, and uses an attention mechanism to fuse weighted neighboring features, so that redundant information can be reduced both before neighboring features are coupled. It can also reduce the semantic differences between multi-highdimensional features and low-dimensional features after fusion. Second, a dual decoding module is constructed to reduce the feature loss in the decoding stage, and a structural loss function is designed to optimize the network for the multi-scale output in the dual decoder. The proposed WFF-Net has been conducted on three datasets, and it shows that the proposed WFF-Net outperforms several existing DL methods in mean intersection of union (NEU-SEG: 85.70%, DAGM 2007: 86.12%, MT defects: 82.72%) and F1-measure (NEU-SEG: 94.11 %, DAGM 2007: 96.32 %, MT defects: 93.90 %).