<![CDATA[This study aims to analyze the impact of image enhancement techniques on welding defect detection performance using a deep learning-based YOLO11L model. The dataset consists of 1392 welding images categorized into four classes: Good, Crack, Porosity, and Bad, with a significant class imbalance. Five image enhancement methods were evaluated, namely Zero-DCE, RETINEX, CLAHE, Supervision, and Gamma Correction, and compared against a no-enhancement baseline. Image quality was assessed using SSIM, and PSNR, while detection performance was evaluated using Precision, Recall, F1-Score, and mAP50. The results show that Gamma Correction achieves the best image quality improvement, with an average SSIM of 0.569, and a PSNR of 18.862 dB. However, contrasting results are observed at the detection stage, where 0.7772 and 0.6969, respectively, for the Gamma Correction-based model while for the baseline model without enhancement outperforms the enhanced model, achieving a mAP50 of 0.7098 and an F1-Score of 0.6965. This finding reveals a paradox where improved visual image quality does not necessarily lead to better object detection performance. This study highlights the importance of end-to-end evaluation in computer vision systems, particularly in industrial inspection applications, and demonstrates that original images, which are closer to the pretrained data distribution, may yield better detection results than heavily enhanced images.]]>
Copyrights © 2026
