Abstract:

In this study, InceptionV3 based Enriched Feature Integration Network (Inc-EFIN) architecture was developed for automatic detection of surface defects. In the proposed architecture, features of all levels of the InceptionV3 network architecture with the same height and width are extracted and combined. As a result of merging, 5 feature maps with different dimensions were obtained. Channel-Based Squeeze and Excitation (CSE) block has been applied to reveal important details in these feature maps. The CSE block strengthens important details by examining the channels in the feature map. In Feature Pyramid Network (FPN) module, information from low-level feature maps containing spatial details were transferred to high-level feature maps containing semantic details. Then, for the final feature map in the proposed architecture, 4 different feature maps obtained at the end of the FPN module were combined using the Feature Integration and Signification (FIS) module. The feature map combined in the FIS module was passed through the Spatial and Channel-based Squeeze and Excitation (SCSE) block, enhancing the spatial and semantic information that may be important for defect detection in the best way. Defect detection results were obtained by using convolution and sigmoid layers in the last layer of the Inc-EFIN architecture. MT, MVTec-Texture, and DAGM datasets were used to calculate the pixel-level defect detection success of the Inc-EFIN architecture. In experimental studies, Inc-EFIN architecture achieved higher performance than the latest technologies in the literature with 77.44% mIoU, 81.2% mIoU and 79.46% mIoU performance results in MT, MVTec-Texture and DAGM datasets, respectively. Keywords: Pixel-Level Surface Defects Detection, Convolutional Neural Network, Squeeze and Excitation Block, Feature Pyramid Networks

Keywords: