Abstract

Purpose of research. Corrosion recognition on metal structures is a serious problem in conducting inspections of industrial facilities. Existing approaches to image analysis use all images to recognize areas damaged by corrosion, which is not suitable for structural analysis, since the percentage of errors in this approach is very large. Under conditions of corrosion prediction throughout the image, errors related to predictive mask not on metal structure are possible. Therefore, it is necessary to delete the results of positive class prediction for areas damaged by corrosion but not placed on metal structure. Therefore, in this work, the authors have developed two-step approach for recognizing corrosion of metal structures, thereby achieving the goal of improving recognition accuracy.Methods. We implement two deep learning models focused on Semantic segmentation (DeepLabv3, BiSeNetV2) for corrosion detection that work better in terms of accuracy and time and require fewer annotated samples compared to other deep models, such as Unet, FCN, Mask-RCNN. A new detection approach to metal areas damaged by corrosion, based on the combination of two convolutional neural networks for more accurate pixel prediction by depth architecture models: DeepLabv3 and BiSeNetV2.Results. Experimental studies have calculated the accuracy and F1 measures using FCN, Unet, Mask-RCNN models as well as the proposed approach. Based on obtained results, it was concluded that proposed approach of combining DeepLabv3 and BiSeNetV2 networks increases accuracy and F1 measure for Unet algorithm by 3%, accuracy by 10% and 2% F1 measure for Mask R-CNN and by 12% accuracy and 4% F1 measure for FCN network. Experimental results and comparisons with real data sets confirm the effectiveness of proposed scheme even for very complex images with many different defects. Productivity was assessed based on data annotated by experts.Conclusion. Analyses of existing solutions in the field of recognition of metal structures damaged by corrosion is described. Shortcomings of existing solutions based either on detection of corrosion sites or on pixel segmentation of full image are identified. A new approach to the recognition of metal areas damaged by corrosion based on the combination of two convolutional neural networks for more accurate pixel prediction of DeepLabv3 and BiSeNetV2 is indroduced. Production is evaluated based on data annotated by Precision and F1-score metrics experts.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call