Research On Single Image Reflection Removal Algorithm Based On Deep Learning

At present,reflection removal has become a research hotspot in the field of artificial intelligence.Machine learning and deep learning can effectively eliminate the reflection effect of images,but still face many difficulties.The reflection layer feature information is complexly coupled with the background layer feature information,and the separation is difficult;the mutual comparison of multiple input images is used to peel off the reflective layer and the clear background layer,but the processing of getting multiple images is time consuming and laborious;Keep highfrequency details of the original information cannot be kept and it leads to lose relevant feature information.Therefore,it is especially important to be able to remove reflection from a single image and better restore visual content in images.Aiming at the problems encountered in the above removal of reflective task,this paper analyzes the background image,the reflection layer image and the characteristic difference of the original input image mixed with reflection,the deep learning algorithm is used for feature extraction,and then classifies two kinds of images,and finally the reflection of the input image is eliminated.The main work is as follows:A single image reflection elimination framework based on residual learning is proposed.The relationship between the input image and the residual image is completely mapped.The corresponding feature information is compared to obtain the difference between the reflection layer and the background layer,and the image is completed to the maximum extent.A deep memory network is proposed,which is analogous to the synaptic structure of the human brain.It associates the previous state with the current state to form a short-term memory mode.Multiple short-term memory modes constitute a long-term memory unit.Experiments prove that both frameworks can be used to effectively remove the reflection effect in the image.The confrontation training is generated by GAN,the deep residual blocks and the deep memory units are designed in the generator,the convolutional layers and the batch normalization layers are designed in the discriminator.By modeling the distribution features of the input image,the antagonism between generator and discriminator is studied are modeled.As a result,the image effect of the output residual image is enhanced,and the effective gradient information is further provided.The whole model is guided in the form of feedback to learn the image features better,and the coherence is more realistic and more in line with human vision.The effect of the image texture details is enhanced.Finally,the objective function of the whole deep learning network is designed as a combination of multi-balance loss functions.The pre-trained VGG-19 model extracts image features,and uses the 1 loss function to measure the perceived loss functions.The content loss function can be measured the characterization difference between the target image and the generated image during the first half of the network training;dealing with the texture details in the image against the loss function;and the gradient loss function minimizing the correlation between the predicted reflection layer and the background layer in the gradient domain.The experimental results show that the PSNR of the image evaluation index reaches 24.84 and the SSIM value is 0.892,which have achieved good results.