Small Vessel Detection Based On Multi-channel CNN In Polarimetric SAR Images

Synthetic Aperture Radar(SAR),with its all-day,all-weather imaging capability,can provide continuous monitoring of the ocean over long periods and has significant advantages in the detection of marine vessel targets,and has received a lot of attention from researchers.Vessel detection methods based on Convolutional Neural Networks(CNN)can automatically extract features to achieve higher accuracy than traditional thresholding methods.However,small vessel targets generally have weaker scattering intensity and fewer pixels,and fewer features in the low-resolution feature maps of conventional multi-scale detection,which are prone to miss detection;in addition,the traditional cascade approach of land shielding by sea-land segmentation followed by vessel detection is complex to implement,time-consuming and does not exploit the intrinsic correlation between the two tasks;polarimetric SAR can provide different polarimetric features of the target,enriching the target information and helping to increase the differentiation of small vessel targets from other interfering targets.However,the multi-channel feature of polarimetric decomposition has not been fully applied to vessel target detection.To address the above problems,considering that multi-channel CNN has better feature extraction and decoding capabilities than conventional CNN,the ability to process sea-land segmentation and vessel detection simultaneously,and the potential to adaptively fuse multi-polarimetric channel SAR data,therefore,this thesis carries out research on polarimetric SAR small vessel detection based on multi-channel CNN.The main work and innovations are:(1)A small vessel detection method based on single-scale and low-level feature aggregation is proposed.To address the problem that small vessels have fewer features in the low-resolution feature maps of conventional multi-scale detection and are prone to missed detection,a single-scale feature map suitable for small vessel detection is selected through feature visualization and comparison experiments,and a semantic enhancement module is designed to compensate for the lack of semantic information on the single-scale feature map.Compared with the benchmark detection algorithm,the method improves the average accuracy by 1.55% and 0.18%on the SAR image small vessel detection dataset and the public SAR ship detection dataset(SSDD),respectively.(2)An integrated network structure based on multi-channel decoding for sea-land segmentation and vessel detection is designed.To address the problems of complex and time-consuming cascading methods of sea-land segmentation and vessel detection without considering the intrinsic correlation between them,multi-channel decoding is used to synchronize sea-land segmentation and target detection in one network,which shares the global semantic information extracted from sea-land segmentation during vessel detection,improves the discrimination ability of small vessel detection model for land false alarms and reduces missed detections.(3)A small vessel detection method based on an adaptive fusion of multi-channel polarimetric features is proposed.To address the problem that the multi-channel information of polarimetric decomposition has feature redundancy and is not fully utilized,this thesis uses multi-channel encoding to achieve adaptive fusion of polarimetric features,which automatically extracts features that help small vessel detection while reducing feature redundancy.Compared to the detection algorithm using the amplitude map,the average precision of small vessel detection is improved by 3.83% and7.9% on dual-polarimetric and quad-polarimetric data,respectively.