Image Restoration And Saliency Detection Based On Deep Learning

Image editing and image understanding are currently widely used techniques.Image editing technique is used to assist in taking and post-creating photos;image understanding technique is used to classify photos,retrieve images,and provide environmental understanding capabilities for autonomous driving applications.However,the quality of the photos may not be high due to environmental factors such as haze and shadows,which may not only lead to unsatisfactory image editing results,but also seriously reduce the accuracy of image understanding.Therefore,image restoration technique is required to restore the images to improve their quality;in addition,the background of the photo taken is often very complex,and it is very important for image editing and image understanding to capture the main(salient)content of the image,so saliency detection technique is needed to extract the salient region of the image.In order to improve the reliability and performance of image editing and image understanding,this paper studies haze removal and shadow removal of image restoration and saliency detection,and proposes corresponding algorithms based on deep learning:1.To remove haze in an image,so as to improve the performance of image editing and image understanding,existing methods need to first estimate the transmission map corresponding to the input hazy image,but in some usual cases,it is difficult to predict the transmission accurately,which in turn leads to over-dehazing or under-dehazing.This paper proposes to use four additional layering-separation models to supplement the existing atmospheric scattering layer separation model,and combines the attention mechanism to fuse multiple layer-separation results.The experimental results show that the proposed method can achieve the best dehazing results.2.To remove shadow in an image,so as to improve the performance of image editing and image understanding,it is necessary to detect the shadow region first,and then restore the underlying true pixel value of the shadow area.Considering that the multilevel features of the deep convolutional network have context information with different receptive fields,this paper proposes a bi-directional feature pyramid network combined with recurrent attention residual modules to produce accurate shadow detection results.Then this paper uses a deep residual network with dilated convolution and channel attention mechanism to restore the shadow area to remove the shadow.Experiments show that this method achieves the best shadow detection and removal results.3.In order to support image editing applications and improve the efficiency and accuracy of image understanding,saliency detection needs to be applied to find and segment salient objects in an image.Existing methods often use the low-level features of deep convolutional networks to supplement detailed information to high-level features.Due to the weak semantic discrimination of low-level features,this approach introduces noise and misleads the saliency prediction.This paper proposes to use the low-level and high-level features alternatively in multiple refinement stages,and combine the residual refinement blocks to refine the saliency detection results,which can achieve the best saliency detection performance.