Image Restoration Based On Attention Mechanism Convolutional Neural Networks

In the process of acquiring images,the acquired images are inevitably degraded in quality to varying degrees due to the inherent environmental conditions(e.g.lighting conditions,glass reflections,rain,fog).On the one hand,these degraded images are visually unsatisfactory;on the other hand,the noise,occlusion and other information in the degraded images interfere with the subsequent image analysis and processing.Image restoration aims to recover high quality images from degraded images and plays an important role in the field of image processing.Convolutional neural networks have powerful representation learning,and it are widely used in various fields.The attention mechanism selectively locates the important feature information of the image by simulating the characteristics of human vision,which can effectively enhance the feature representation capability of neural networks while reducing the computational resource consumption.Aiming at the problem of image restoration,this paper takes the image mixture model as the base,combines convolutional neural network and attention mechanism to study the restoration task of degraded images in different scenes,and improves the image restoration method in a targeted way.The main research work of this paper is as follows:(1)For the problem of reflection image restoration,this paper designs a reflection image restoration network by combining the reflection image mixing model,which is implemented by using a joint attention mechanism.The network is made up of an encoder and three corresponding decoders,where each of decoder performs a divergent forecasting task.The inclusion of a joint attention mechanism in the transport layer of the prediction branch,which intends to further enhance the transport layer features of the images,direct the decoder to perform feature picking decoding,thus far boosting the prediction efficiency of the network.Considering the gradient property of reflection images,this paper introduces a recovery loss composed of both the gradient loss and the pixel loss in this network to be monitored to train network.Experimental results on different types of reflection images show that the proposed method has good restoration performance in both qualitative and quantitative analysis.(2)For the remote sensing image defogging problem,this paper designs a remote sensing image defogging network based on dark channel attention with deep learning driven by dark channel prior knowledge.The network adopts the parallel architecture of attention flow and dark channel prior constraint flow.In the attention flow,the image extracts feature information through encoder-decoder and combines channel spatial attention structure enhancement features.In dark channel prior constraint flow,the network performance is improved using dark channel priori loss to pass feature information for the attention flow and enhance the ability of the attention flow to learn image features.Meanwhile,considering the importance of edge information in images,an edge loss is introduced to supervise the trained network.Experiments were carried out on real hazy datasets and synthetic hazy datasets.The results show that the method proposed in this paper can effectively remove the non-uniformly distributed haze as well as the haze in the natural acquisition remote sensing images.And the method proposed in this paper has achieved good results in both qualitative analysis and quantitative indicators.