Evaluation Metrics in Digital Image Processing: A Task-Oriented Review and Practical Selection Framework

Abstract

The rapid development of digital image processing and machine learning has increased the need for reliable and task-appropriate evaluation methods. However, the wide range of available metrics, each with different assumptions, strengths, and limitations, makes metric selection challenging across diverse image-processing tasks. This article presents a task-oriented review of commonly used evaluation metrics and proposes a practical framework for selecting appropriate metrics across five major categories: image quality assessment, image classification, image segmentation, image restoration, and object detection. Representative metrics are reviewed based on their mathematical basis, interpretation, applications, advantages, and limitations. Particular attention is given to pixel-wise measures, including MAE, MSE, and RMSE, perceptual and structural measures such as PSNR, SSIM, MS-SSIM, FSIM, VIF, and LPIPS, and classification and detection measures including accuracy, precision, recall, F1-score, ROC-AUC, confusion matrices, and mean average precision. Overlap-based metrics such as Dice and IoU are also examined for segmentation. The review emphasizes that no single metric adequately captures all dimensions of image-processing performance. Therefore, metric selection should consider task characteristics, data properties, reference availability, and perceptual requirements, with complementary metrics used where appropriate. The proposed framework aims to support researchers and practitioners in selecting appropriate and interpretable evaluation metrics for digital image processing and machine learning applications.