Research On Fine-grained Detection Algorithm Of Objects In High-resolution Visible Light Remote Sensing Images

In recent years,deep learning methods represented by CNN have been extensively studied.Deep learning methods can independently learn and mine potential,high-level,and abstract features in images.This learned feature has better discriminative performance than traditional artificially designed features.At this stage,the target recognition method based on deep learning has become the mainstream of image understanding and analysis.In the field of computer vision,there has been a large amount of work showing that using deep learning for refined target recognition can achieve satisfactory results.At present,research in this direction has become a research topic.In the field of remote sensing,the refined classification and recognition of ground objects in high-resolution visible light remote sensing images have broad application prospects in both civilian and military fields.However,in this field,the research on the fine classification and recognition of objects in high-resolution visible light images is still in the preliminary stage,and there are few public reports on fine classification and recognition.The research on the fine recognition and classification of features in high-resolution remote sensing images is facing severe challenges.First,in terms of high-resolution remote sensing image data,there is a lack of public data sets;second,in terms of recognition methods,the difference between remote sensing images and natural scenery images is significant.Remote sensing images usually capture the top information of geospatial objects,while natural scene images usually capture the outline information of the objects.Therefore,the target detector learned from natural scene images is not easy to apply to remote sensing images.Third,it comes from images with high spatial resolution.Similar remote sensing features are diverse and have different local appearance structures.The improvement of the resolution of high spatial resolution remote sensing images,on the one hand,makes the local appearance structure of the features clearly visible,which is convenient for image representation and observation.On the other hand,it also enlarges the intra-class differences of similar features and increases the intra-class variance.Instead,it creates difficulties and challenges for refined recognition.Fourth,it comes from practical engineering problems.At present,the military and civilian fields have an urgent need for the fine classification and recognition of remote sensing image objects in specific scenarios.However,considering the challenges of the large amount of remote sensing imaging data,imaging quality,image noise and other factors,how to integrate the recognition method and engineering The combination of applications is also an engineering problem that needs to be solved urgently.In order to solve the above-mentioned problems in the fine recognition and classification of ground objects in high-resolution remote sensing images,this paper has made preliminary research and attempts in the fine detection of ground objects in high-resolution visible light remote sensing images,and carried out research and practical work in related fields.,Established a high-resolution visible light remote sensing image aircraft fine-grained data set,we named it DFAD,and proposed an improved refined rotating frame method,and used mainstream deep learning targets on the DFAD data set and other benchmark data sets The detection methods have been compared,and good experimental results have been achieved.The work of this article can be summarized into the following aspects:(1)Created a high-resolution visible light remote sensing image aircraft fine-grained data set(DFAD),including 18 sub-categories(1 passenger aircraft,17 military aircraft),filling the current fine-grained high-resolution visible light remote sensing image The field of granular recognition lacks the problem of public data sets.(2)A benchmark experiment was performed on the DFAD data set,which verified that the data set is challenging and can perform performance evaluation experiments of fine-grained recognition algorithms.(3)Designed a refining algorithm for the rotating detection frame,which can optimize the current horizontal detection frame to obtain an improved rotation detection frame.The algorithm performance experiment was carried out on the three benchmark data sets of DFAD,DOTA,and HRSC2016.Finally,it is proved that the refining algorithm of the rotating detection frame proposed by us has the speed advantage of the single-order detector,and the detection accuracy is not inferior to the two-order detector.