Research On Character Image Coloring Based On Deep Learning

Image coloring has been widely used in computer graphics and has became a research hotspot in the field of image processing.With the rapid development of deep learning in recent years,image coloring technology has been further improved,but the existing methods do not have a separate study on the image of the person,and there is a phenomenon that the coloring effect is single and the color is not true.Based on the convolutional neural network for the reasonable coloring of character images,this paper focuses on the realism of coloring effects,the guidance of human interaction and the diversity of coloring.The core work and research results of the paper mainly include the following aspects:(1)Achieve accurate segmentation of character images.In order to study the images of the character,this paper selects about 5,000 images of the character and the plant from the Image Net dataset,and produces a small dataset containing only the characters and the backgrounds.In order to obtain the image segmentation results,this paper improves the U-net network,performs three downsampling and three upsampling,increases the dilated convolution,and uses the sigmiod activation function instead of the Re LU(The Rectified Linear Unit)activation function,and the batch normalization(BN)is placed before the activation function.Reduce the training time of the network and achieve higher quality segmentation results.(2)Achieve reasonable coloring of the image scene.Due to the lack of realism in scene coloring,this paper uses a global hint network method,and uses local mean method to reduce the color resolution of the real image,and converts it into HSV(Hue,Saturation,Value)color space and calculates the global histogram,then Uses the Euclidean distance to match the best image.Finally,the fusion layer is used to connect the global hint network to the main coloring network to achieve the purpose of reasonable coloring of the image scene.(3)Generate a variety of reasonable coloring effects in real time.The advantage of the end-to-end learning framework is that it can adapt to different types of user input and generate coloring effects in real time.On this basis,this paper adds a local hint network,obtains the ab color gamut of CIE L*a*b*quantization,and uses K-means clustering to find the color distribution pattern in the image,and then provides discrete color suggestions for the grayscale image.Then,the Gaussian sampling is used to set the user control points,and various effects of image coloring are realized through manual interaction.Finally,this paper designs and implements the coloring system of black and white character images,and shows the experimental results and analysis.The results show that the method described in this paper can meet the needs of the realism and diversity of the coloring effect of the characters,and has important guiding significance for the research of image coloring.