Semantic Segmentation of Defects in Infrastructures through Multi-modal Images

The accurate segmentation and detection of defects in infrared and visible images are critical for non-destructive testing applications, however those steps are often excluded by limited annotated training data. This paper presents an innovative approach for the segmentation and detection tasks into a unified framework. The proposed method introduces and tests a novel framework tailored to the domain of infrared and visible imaging. This framework eliminates the need for annotated defect data during training, enabling models to adapt to real-world scenarios where annotations are scarce. To enhance the accuracy of segmentation and detection, it employs super-pixel segmentation, following by texture analysis.