Evaluation of the algorithm and method qualitys for portrait photographs colorization

A comprehensive experimental study was conducted to compare the quality of automatic colorization algorithms for portrait photographs. Four different approaches have been analyzed: the classical K-nearest neighbors method and three neural network models – ECCV16, Siggraph17, as well as the modern Dual Decoder Colorization architecture. The quality of colorization was studied on a specialized dataset of celebrities' faces, consisting of 400 test images in JPG/PNG format. Key differences in the architecture and operating principles of the methods under consideration, including the use of convolutional networks, transformers, and traditional machine learning algorithms, have been identified. Objective evaluation criteria were determined using three standardized metrics: peak signal-to-noise ratio, structural similarity index and color difference metric. Quantitative test results have been established demonstrating the leadership of the Dual Decoder Colorization model in terms of structural similarity (SSIM = 0.929) and visual quality. The results of a visual comparative analysis are presented, vividly illustrating the semantic consistency and naturalness of color rendering of various methods. Conclusions are drawn about the balance of modern neural network solutions Dual Decoder Colorization and Siggraph17, as well as about the limited effectiveness of classical K-nearest neighbors and ECCV16 approaches for high-quality colorization tasks. Practical recommendations have been developed for choosing a model depending on the requirements for saturation and accuracy of color reproduction. The expediency of using modern architectures to achieve realistic and visually appealing results of automatic colorization of portrait images is substantiated.