Automatic Detection Method Of Multi-class Main Targets In Remote Sensing Image Of Port Area

With the development of remote sensing image observation technology in recent years,the resolution of remote sensing images has gradually increased.Nowadays,object detection in high-resolution remote sensing images has become a research point.As a common scene in remote sensing images,ports are an important part of human maritime traffic activities.With the development of the economy,the trade and transport needs of the port make the management and monitoring of the port a big challenge,and it can have great dynamic changes.Therefore,how to use highresolution optical remote sensing images to detect various types of targets in the port area has become an urgent need.Currently,ocean scenes in high-resolution remote sensing images of harbor areas tend to cause false alarms and other interferences with target detection results;at the same time,black scenes have complex backgrounds,difficult feature extraction,large multi-scale differences.traditional detection methods cannot perform efficient feature extraction and detection processing,and it is difficult to accurately detect highly organized targets in the scene.In view of the abovementioned problems,this research builds a method for automatic detection of multiple main targets in port area remote sensing images,which mainly includes two aspects:first,to propose a fast preprocessing module for sea and land segmentation based on depth perception,and second.Second,to propose multi-scale fusion and a Target detection algorithm based on feature enhancement.The main content is as follows:(1)In view of the complex scene of marine area in high-resolution optical remote sensing images,some target features of islands,ships and port areas are similar,which is easy to produce a large number of false alarm interference and affect the detection accuracy.In this paper,a fast preprocessing module of sea-land segmentation based on depth perception is constructed.The remote sensing image sea and land area is segmented by the fast segmentation preprocessing module,so as to eliminate the sea area in the image,reduce the interference of the sea area to the multi-target detection of the port.Aiming at the problems of large differences in multi-scale features of remote sensing images and blurred boundaries between sea and land in remote sensing images of port areas,we first propose an RS-Deep Lab segmentation framework to obtain shallow supertransverse,deep features and mid-level features for feature fusion to improve the robustness of the model.Then,we propose a densely connected multi-scale feature extraction enhancement module to expand the receptive field,overcome the problem of low segmentation accuracy at different scales,and solve the problem of blurred boundaries to improve the segmentation accuracy of the segmentation network.In addition,we propose the RS-Rep VGG network as the backbone network of the segmentation preprocessing module.The structural reparameterization is used to optimize the reasoning speed of the network,and the network structure is optimized to improve the timeliness of the sea-land segmentation preprocessing module.It provides a basis for subsequent multi-target detection in port areas.Second,to address the problem of large differences in the scale of vehicle targets in remote sensing images with different resolutions,this thesis proposes that a method based on weighted bidirectional feature pyramid fusion is used in the feature fusion stage.By constructing top-down and bottom-up bi-directional channels in addition to forward propagation,the feature information from different scales of the backbone network is fused,thus effectively solving the problem of large scale differences of vehicle targets in remote sensing images with different resolutions.(2)In order to accurately detect the results of the above segmentation preprocessing module,this paper proposes a detection algorithm based on multi-scale fusion and feature enhancement.This paper first selects the current mainstream detection network Yolov5 as the basic framework.In view of the complex scene of high-resolution remote sensing image.The difficulty of feature extraction.The small difference of similar features,and the dense arrangement of targets.This paper first proposes RS-Darknet feature extraction backbone network.The RS-CSP module is used to improve the feature extraction ability of the backbone network,and then the MLFPN feature fusion module is used in the feature fusion part to perform cross-scale connection of the extracted feature layers of different scales to enhance the multi-scale feature expression ability and the robustness of the network to multi-scale targets.At the same time,we construct a dual attention feature enhancement module,which mainly includes position attention and channel attention.Through the dual attention module,the expression ability of boundary features is enhanced,and the feature information of densely arranged targets in the port area is enhanced,and the difference between inter-class features is increased,so as to solve the problems of dense arrangement of targets,inter-class feature coupling and boundary blur.In this paper,the corresponding comparative experiments and ablation experiments are carried out on each module and network framework proposed above in the large open remote sensing data set XView data set.The feasibility of the method studied in this paper is proved by experiments.