ELF-CandyGAN: A Candy Color Coloring Method for Image Local Feature Enhancement

Candy color is a new phenomenon in the field of photography, and its high brightness, low saturation, and low contrast bring a unique color experience to the world. The CandyCycleGAN network is good at realizing the transformation from ordinary color to candy color. Still, there will be problems as some of the image details are not dealt with properly, so to solve the above problems, this paper designs a candy based on the local feature enhancement of the image color coloring method (ELF-CandyGAN). Based on the generator of the U-Net network, a color learning module is designed in the downsampling process to learn the distribution, relationship, and features of candy color, which helps the network to better learn and understand the color information and keep the naturalness and authenticity of the color, and at the same time, a unique jump connection is designed to add the results in the upper layer convolution as part of the results in the lower layer convolution; secondly, a global context module is introduced in the color learning module. To ensure that the chromaticity values learned during the entire network training process remain within the candy color range, a global context module is introduced. This module also reduces computational complexity and accelerates network training. Subsequently, a feature enhancement module is designed, which introduces additional enhancement operations to enable deeper mining and processing of network features, thereby improving the performance and effectiveness of the network in the coloring task. This module also helps to enhance and restore the detail information lost during the downsampling process. Furthermore, a dual discriminator network based on PatchGAN is constructed. The first discriminator, D1, adopts a multi-scale discriminative structure to guide the generator toward producing richer image details. The second discriminator, D2, is designed to compute the structural similarity between the generated image and the original input image, encouraging the generator to produce structurally consistent outputs. Finally, a feature structure loss function is proposed to impose constraints on the structural similarity between the generated and input images, ensuring that the generated images retain more original detail features and exhibit higher realism.