Distortion Measurement Metric for Human Image Refinement and Evaluation Using Distorted Image Datasets

This paper introduces novel evaluation metrics for quantifying distortion in human image retouching and presents a COCO-based retouched image dataset to validate their effectiveness. The dataset is constructed using human images from the COCO dataset, which provides a large-scale and diverse collection of images, including a variety of facial structures, poses, and lighting conditions, making it well-suited for human-centered retouching research. Despite extensive research into the perceptual assessment of visual discomfort induced by distorted images, a significant gap exists in the availability of quantifiable evaluation metrics and dedicated distortion datasets. To address this issue, we developed two distinct evaluation metrics and generated an optimal dataset tailored for this purpose. The two-evaluation metrics proposed in this study are the Distorted Line Similarity (DLS) metric, which uses edge detection, and the Point-to-Point (P2P) metric, which leverages corner detection. In contrast to traditional image quality assessment metrics such as PSNR and SSIM, these new metrics quantify the changes in lines, curves, and feature points within an image, providing meaningful results when assessing image distortion. Furthermore, the proposed framework produces results with reduced distortion compared to existing retouching applications. Using the proposed metrics, we demonstrate that the images generated by the proposed framework exhibit significantly less distortion than the original distorted images. The demand for retouching images in which the main object is a person increases. Excessive distortions can be quantified and detected in contexts such as immigration procedures using the proposed metrics. By publicly releasing the COCO-based retouched image dataset, which includes the original images, distorted images, and images with minimized distortions used in our experiments, we aim to demonstrate the quality of our dataset and contribute to the field of image distortion minimization research.