Research On Efficient Colorization Based On Deep Learning

The color of images plays a crucial role in the proper functioning of various computer vision systems in everyday life,such as access control systems,automatic driving and medical diagnostic systems.Image colorization aims to convert a given grayscale image into a complete and believable color image,which is a very challenging task in image processing and computer vision.With the development of deep learning and big data,the end-to-end training model has greatly reduced the number of human interventions in image colorization,so building deep learning-based image colorization models has become the current trend in this field.For better results,it is often necessary to design complex network structures,which limits the use of colorization models in real-life applications.How to reduce the number of parameters and the computational effort of the model as much as possible while maintaining the coloring results is currently the research focus in the colorization tasks.Therefore,this paper designs a deep learning-based lightweight image colorization model with the goal of improving the comprehensive performance of colorization methods,which mainly includes the following:1.An Image Colorization model based on a stacked multi-scale attention is proposed in this paper.First,we design a Stacked Multi-Scale Attention(SMSA)mechanism to obtain distinctive and specific features from multiple scales for various images.Since SMSA integrates multiple information from multiple feature scales instead of a single scale,it is more comprehensive and reasonable than existing attention mechanisms.Moreover,a foreground loss is proposed to work with the existing smooth L1 Loss to regularize the details of the images at both pixel and object levels.Compared to other object-based colorization methods,foreground loss does not require additional parameters and computational effort in the inference phase,which greatly reduces the speed of model inference.2.To reduce parameters and computation of the colorization model,a global-local attention-based channel pruning algorithm is designed in this paper.First,we proposed efficient global-local channel attention mechanism,it shift long-range features to the neighborhood by the shuffle-reorder operation,which enables convolutions of small kernel size to capture long-range dependencies.Secondly,we design a channel pruning algorithm based on the proposed attention mechanism,which can adaptively obtain the individual weights of the pruned channels,avoiding the network to eliminate important feature channels and convolution weights.We demonstrate the efficiency and effectiveness of the proposed lightweight colorization model on ImageNet and CIFAR-10 benchmark datasets.