Research On Single Image Rain Removal Algorithm Based On Meta-learning

Nowadays,computer vision technology involves all aspects of people's lives,such as industrial inspection,medical treatment,traffic monitoring,and autonomous driving systems.When taking photos in real life,there are usually many interference factors,these factors will make the image very blurred,and then affect a series of subsequent computer image processing technology.Therefore,how to solve the problem of image quality degradation caused by various interference factors has abstracted more and more attraction in the field of computer vision,which is of great significance to the theoretical research and practical application of image processing technology.As a common weather phenomenon in rainy days,when photographing in rainy days,due to the influence of rain streaks,a lot of image information is often lost,and the image quality becomes very poor.Therefore,we focused on how to remove rain streaks in a single image.There are many studies on image rain removal tasks.Most of them are based on deep learning technology which rely on big data sets used during neural network training.However,it is difficult to obtain a pair of rainy and clear images in the same scene.Therefore,the training data set usually uses artificially synthesized data sets,and the rain image obtained by artificial simulation cannot fully show the rain image of the real scene.And some of them didn't fully consider the more complicated and changeable rainy conditions in real life,and performs poorly when processing rainy images under complex scenarios.In order to solve the above problems,we deeply analyzed the image processing technology based on deep learning.In this work,we chose to start from the perspective of improving the generalization ability of the network by integrating meta-learning and deep learning technology,and proposed a new research direction for single-image rain removal.The specific research contents are summarized as follows:First of all,this paper proposes a training data set that adapts to the meta-learning training method.The meta-training set and the meta-test set are constructed in the way of N-way and K-shot.The meta-training set is divided into the support set and the query set.In each training process,different training rainy images are randomly selected to realize the learning process of rain streaks.Secondly,based on the network of meta-learning,this paper proposes a deeper and more complex convolution neural network,which consists of a 20-layer convolution layers,in which in order to ensure the size of the output of the last layer of convolution layer and the original input image are the same,there are no pooling layers after all convolution layers.The proposed simple convolution network is more suitable for the image feature extraction,and can learn more advanced image features.Finally,the experiments on synthetic data sets and real data sets showed that the algorithm proposed in this paper has a significant improvement in objective values and subjective vision,which provides a guarantee for the subsequent application of computer vision systems.At the same time,this paper also proves the different improvement effects of different modules on the network through ablation experiments.In addition,this paper further extends the proposed algorithm to image defogging,image snow removal and graphics super-resolution tasks.The experimental analysis shows that the algorithm in this paper has a strong generalization ability and can be extended to more image denoising tasks.