Research On Image Color Rendering Algorithm Based On Generative Adversarial Networks

Image color rendering refers to the expression of grayscale pixels in the target image with a certain color to improve the visual effect and highlight useful information.It is one of the important image enhancement technologies in the field of image processing,and has been widely used in product appearance color design,augmented reality,digital entertainment and other fields.At present,common color rendering methods are faced with details such as color boundary crossing and fuzzy boundary.The traditional color rendering method has the limitation of high labor cost and high reference image quality.Deep learning-based color rendering methods have some problems,such as feature loss caused by upsampling,difficulty in model pre-training,gradient disappearance or explosion,under-fitting or over-fitting,and poor robustness.To solve the above problems,this thesis studies the robustness and generalization of the algorithm,the stability of generative adversarial networks,and the lightweight and high performance of the algorithm respectively.It mainly includes the following contents:(1)Aiming at the problems of poor robustness and generalization of color rendering algorithm in low-quality images,an improved pix2pix model method based on Gabor-filter is studied.Firstly,Gabor filter is used to extract multi-scale and multi-direction texture feature map,and the image is preprocessed.Secondly,based on pix2pix model,the least square loss function and gradient penalty term are used to ensure the diversity of generated samples and stabilize the network model training.Finally,the generator is used to render the robust image,which reduces the influence of light change and noise on the image to a certain extent.(2)To solve the problem of poor stability of generative adversarial networks,a Hinge-Cross-Entropy GAN model is studied.Firstly,the self-attention mechanism is improved to effectively capture feature dependence at a distance.Secondly,the Hinge-Cross-Entropy loss function’is designed to strengthen the training effect,so that the loss is always in the optimal state,which is used to stabilize the model.Finally,a Hinge-Cross-Entropy GAN is proposed to achieve automatic image rendering on DIV2K and COCO datasets.Experiments show that the rendering quality and effect are improved.(3)Aiming at the difficulty in balancing the lightweight and rendering accuracy of color rendering models,a frequence channel attention GAN model is studied.Firstly,global average pooling is the lowest frequency domain component of discrete cosine transform.In order to integrate the remaining frequency domain components into channel attention mechanism,a frequency channel attention mechanism is designed.It not only reduces the number of parameters and the amount of computation,but also better captures the rich input mode information.Secondly,frequency channel attention GAN is proposed by combining the frequency channel attention mechanism of U-Net network,which improves the model performance and reduce the model complexity.Finally,the method is implemented in Jittor framework,which saves the cost of computer resources compared with PyTorch framework.