Research On Single Image Deraining Based On Semi-supervised Learning

As the main form of visual information,a single image is widely used in urban traffic,security monitoring,network media,etc.,and is closely related to people’s lives.When a camera lens shoots the outside scene into a single image,it needs to open the shutter,increase the aperture,activate the CMOS sensor and convert analog to digital;The imaging process will produce noise inevitably,and rainy weather will further re-duce the image visual effect.Especially in the field of autonomous driving,noise in the foreground will obscure the moving target in the background,which is not con-ducive to tracking the target;Therefore,the single image deraining algorithm can be applied to multiple high-level vision tasks such as image classification,detection,and semantic segmentation.As a sub-field of low-level image processing,it has been deeply followed by scholars in the industry.Compared with traditional image processing al-gorithms,deraining algorithm with deep learning can achieve excellent results on sim-ulation datasets,but on real datasets is poor,such as background texture oversmoothly or rain streaks removed insufficiently.To solve this problem,how to adopt real rainy images as input for model training is the key problem to be solved.The groundtruth corresponding to a real rainy image is difficult to obtain;So,the Euclidean distance cannot be used to train the model.Therefore,our algorithm adopts a semi-supervised learning strategy,which includes the following two algorithms.（1）Semi-supervised single image deraining based on knowledge transfer;His-togram statistics are performed on the pixels of rain streaks between Rain1400 simula-tion dataset and the real rainy dataset,which the statistical results show similar distri-bution.According to this phenomenon,deraining problem can be transformed into the fitting problem of solving the real rain layer and the simulated rain layer.Let the real rain streak feature as a knowledge inducement to train the network to correct the noise distribution of the simulated rain streak layer;The characteristics of the real rain streak and the simulated rain streak are converted into high-dimensional coding,and their rel-ative entropy is calculated;According to the pixel dark channel theory proposed by previous scholars,the dark channel value of a rainy image is greater than 0 while of a clear image is equal to 0.Therefore,the dark channel loss function of a real rainy image and a simulated rainy image can be calculated,and its difference can be minimized in the iterative optimization process;Our model can learn the pixel distribution of the real rain streak,which improve the robustness of the model to deal with real rain image.（2）Semi-supervised single image deraining based on wavelet transform with dis-criminative learning;Different from weakening the difference in rain streak pattern dis-tribution between the real rainy image and the simulated rainy image,our algorithm adopt wavelet transform to decompose a rainy image into three sub-band maps,such as f_HH,f_HLand f_LH,calculate the discriminant loss of predicted background and un-paired clear image in three sub-band to enhance the robustness of the network model in predicting the background image.