Research On Image Reflection Removal Algorithm

When pictures are taken through transparent glass,the captured image often contains the reflection of the scene on the same side of the camera.The presence of reflection will cause visual interference and reduce the image quality.In computer vision tasks,it is very important to remove reflections in images taken through glass.It is of great significance in improving image quality and preprocessing images.It is an ill-posed problem to decompose the transmission image of the target scene and the reflection image of the scene on the same side of the camera from a mixed image with reflection.Therefore,it is difficult to remove the reflection in the mixed image and needs to be based on reasonable assumptions prior to design the reflection removal method.Research has been carried out on the problem of reflection removal in images,and the main research contents are as follows:1.Aiming at the traditional image reflection removal algorithm to remove the reflection of the whole image,which causes the problem of pixel separation error in the non-reflective region,a new image reflection removal algorithm based on the reflection region is proposed.Firstly,according to the characteristic that the gradient of the reflection region is small,the initial reflection region marker map is obtained by gradient image.Then,the more accurate reflection region marker map is obtained by using both the transmission region marker map and the maximum chromaticity difference binary map.Finally,the reflection of the reflection region is suppressed according to the reflection region marker map.Experimental results show that the algorithm can remove reflections better,and has a better effect in removing reflections and retaining the image details of the transmission layer.2.A new mixed image model is proposed.Based on the mixed image model,the convolutional LSTM network structure proposed by other literature is used to remove the reflection in the image.First,through statistics,it is found that the brightness of the mixed image obtained by the existing mixed image calculation formula is higher than that of the actual reflected image.Therefore,a parameter is designed to adjust the brightness of the mixed image,and a new calculation model of the mixed image is proposed.Then,based on the newly proposed mixed image model,the image reflection removal method based on the convolutional LSTM network structure proposed in other literature is used to remove the image reflection.Finally,training and testing are carried out on three real natural image datasets.Experiments show that the reflection removal method based on the new mixed image model has a good reflection removal effect on various scene reflections,including strong reflections.From the perspective of qualitative and quantitative aspects,it is better than other reflection removal methods based on neural networks.