Research On Dim And Small Target Recognition Technology In Earth Background

Deep learning technology has developed rapidly in the field of segmentation and detection.In the airborne earth observation and recognition target scene,the weak and small target recognition system equipped on the missile-borne or airborne platform has the advantages of wide application range and strong flexibility,and has wide dual-use value for military and civilian use.The existing recognition technology faces the problem of low accuracy of target segmentation and classification,low detection and recognition rate when the dim and small target in the imaging system has limited pixels and is difficult to distinguish from the ground background.Relying on the project of Aviation Science Foundation,the paper carries out the research on the recognition technology of small and weak targets in large background according to the demand of photoelectric observation system.A target recognition technology based on background segmentation and small target cascade is proposed.The feasibility of the technology and the project index can be achieved are verified by experiments.At the same time,based on the existing domestic Huawei Atlas200 DK platform,the algorithm acceleration technology for embedded systems is explored.The main research contents of this paper are as follows:(1)A semantic segmentation model of small and weak targets in ground background is studied.Firstly,A Data set of visible light images with terrestrial background was constructed and enhanced.Secondly,based on the classic Deeplabv3 plus model,a semantic segmentation model of small weak target in ground background is constructed by cascading the backbone network part processed by the attention mechanism proposed in the paper,the ASPP-6 structure of multi-scale small weak target features and the segmentation residual network structure in turn.Then,the model was pre-trained on the public dataset Pascal VOC2007 and retrained on the self-made semantic segmentation training set.Finally,the classification accuracy of the rough segmentation effect between the ground background and the small target was tested based on a validation set.(2)A cascaded detection technology of dim and small target in geodetic background is studied.On the one hand,based on the self-made target detection training set,a fine-tuned yolov7 small target detection model is trained by using a mixed precision method,and the weak and small target recognition model under the earth background is obtained,and the target recognition performance of the model is tested on experiments.On the other hand,aiming at the problem of low coverage and low detection rate of target segmentation or detection network,a cascade detection technology was proposed for dim and small target based on segmentation and detection network for special terrestrial scenes.It can not only coarse segment the dim and small target in the image,but also further detect the location and category of the dim and small target in the ground background.The results show that the cascade detection technology has a good performance on the dim and small target recognition,which effectively solves the problem of low target segmentation rate and missing detection.(3)An application technical scheme of the small target segmentation model with domestic embedded platform is studied.Based on the existing domestic embedded platform(such as Huawei Atlas 200 DK developer kit),a embedded application system of weak and small target segmentation technology for air-to-ground observation was developed.The transplantation of segmentation model and optimization acceleration technology were carried out,and the experimental results of a single process taking 129 ms and a running speed of 10 Hz were obtained.The stability and feasibility of the embedded technology solution are verified.