Remote Sensing Image Dehaze And Ship Detection Based On Deep Learning

Optical remote sensing has developed rapidly in recent years,providing people with richer detailed information in high-altitude perspectives.Among them,the processing tasks for ship targets have become the hotspot of current research because of its unique target characteristics and wide application background.Combined with cutting-edge deep learning theory,this paper starts with the workflow of ship detection to carry out research on enhanced dehazing and location recognition of remote sensing images.The main work of this paper is summarized as follows:(1)A dual weak supervised learning theory is proposed to generate a method for dehazing the remote sensing image.Aiming at the dealing with the shortcomings of the current deep learning dehazing method relying on paired data sets,a dehazing network based on Generative Adversarial Network is proposed.The network directly uses unpaired real hazeno-haze images to learn the haze-to-no-haze implicit mapping conversion relationship by adding a loop-aware loss-encoder-decoder network,while setting its dual network learning to haze-free.The hazy mapping transformation relationship,the two networks compete against each other,through training to get a generator that can directly output haze-free images from haze pictures.The method is tested on artificially synthesized haze images and real remote sensing haze images,and image quality has been improved in related evaluation indicators such as peak signal-to-noise ratio,and the generated images are closer to real hazefree images,which enhances the visual effect.(2)A ship detection method based on improved feature coding part and region of interest sampling method is proposed.In the feature coding part of the network,the feature map size of the last three layers of the dilated convolution hold feature extraction structure is used,to make sure receptive field is constant,while the low layer position information is not scaled and performace in the detection of multi-scale ship targets has improved;then the original direct pooling method is replaced with bilinear interpolation alignment in the sampling process of the region of interest,avoiding candidates due to two integer quantization Frame offset;finally,a non-linear non-maximum suppression method is used to reduce false deletions for overlapping preselected boxes.Experiments show that the proposed method improves the accuracy and recall rate for the ship target in the horizontal rectangular contour frame marking method.(3)A ship detection method based on segmentation model and anchor-free border generation is proposed.Different from the mainstream horizontal rectangular outline box labeling method,this section uses a more precise angular contoured rotating contour bounding box,and proposes a method of directly using pixel points to perform arbitrary angle border regression on the segmentation result,and by adding The loss function that constrains the overlapping part of the sample and the non-maximum suppression method based on the category confidence complete the ship target detection marked by the rotating contour box.Experiments on the data sets of the new rotating dataset DOTA and HRSC2016 show that the method is effective for the special task of ship detection,and has achieved significant performance improvement compared with the existing methods.