Context and scale-aware YOLO for welding defect detection

Radiography testing for welding defect detection is an essential inspection procedure to ensure welding quality. However, detecting these defects is a challenging task because they have various size and aspect ratio characteristics and low perceptiveness due to the low luminance and contrast characteristics of the radiography image (RI). To address these difficulties, this paper proposes a twin model-based automatic welding defect detection method to reveal welding defects of various sizes and aspect ratios more accurately. In addition, we propose a new image adjustment technique that is optimized to improve the accuracy of welding defect detection by adaptively adjusting the luminance and contrast of a given RI. The proposed method consists of three steps: preprocessing for defect detection (PDD), context-aware image adjustment (CIA), and scale-aware defect detection (SDD). In the PDD step, we extract the region of interest from a RI based on text detection by removing regions unnecessary for welding defect detection. In the CIA step, we adaptively optimize a given image to improve the detection accuracy by utilizing a differentiable parametric module that performs image enhancement filtering. In the SDD step, we define a twin model that outputs the embeddings of different scales from the adjusted RI to detect the defects with various scales accurately. At the inference stage of the detection model, we ensemble the results using a weighted fusion of the detection results from the twin model to take advantage of the ensemble strategy. The experimental results indicate that the proposed method achieves outstanding detection accuracy compared to the benchmark methods.