Research On Small Scale Object Detection Methods For Astronomical Images

Object detection is one of the research hotpot in computer vision,which forms the foundation of many visual tasks,such as instance segmentation,image annotation,and object tracking.With the development of the Sloan Digital Sky Survey project,more and more astronomical data have been accumulated.Small-scale objects in these data are important materials for us to explore the formation and evolution of the universe.This type of astronomical objects has the characteristics of low brightness,small scale,and carrying less feature information,greatly increasing the difficulty of automatic search and recognition of such targets.In addition,complex background and large and bright objects cause significant interference to the small-scale object detection.In order to reduce interference and improve the detection accuracy of small-scale objects,this article conducts in-depth research on the removing of interference and the small-scale object detection method in astronomical images.The main research content is as follows:(1)The objects with large scale and high brightness in astronomical images greatly interfere with the detection accuracy of small scale objects.To solve this problem,a removing interfering objects and inpainting image method based on Mask GAN is proposed.Firstly,adaptive threshold segmentation,connected domain analysis and contour detection are used to construct an interference target mask;then,a GAN model based on gated convolution was constructed,which combined with the mask of interfering objects to complete the interference removing task.Experiments have shown that this method can effectively remove interfering objects and inpainting image,which lay the foundation for future small-scale object detection methods.(2)The deep learning object detection algorithm does not pay more attention to small objects and not extract enough features.When applied to small-scale objects detection tasks in the astronomical field,the detection accuracy is low.Therefore,a detection method of small-scale astronomical objects based on YOLOv7-4BES is proposed.Firstly,an attention module EBAM combining channel and space is proposed to enhance the model’s attention to objects area.Secondly,a 4-layer bidirectional weighted feature fusion structure 4B-Bi FPN is constructed to fully fuse more multi-scale features.Then,a shallow feature enhancement module SFM is proposed to solve the problem that the receptive field of shallow feature is small and lacks sufficient context information.Finally,the Focal loss function is introduced to dynamically reduce the weight of easily distinguishable samples and focus on hard distinguishable samples.The experimental results show that the improved model outperforms the current classic object detection algorithms in terms of evaluation indicators.(3)Some small-scale astronomical objects have the characteristic of low surface brightness,which increases the difficulty of automatically and effectively searching for such targets.Therefore,a low brightness small target detection method based on mask mechanism and target cross recognition is proposed.Firstly,a masking mechanism was designed,which combined image segmentation and source detection methods to remove the majority of non detected targets from the image,resulting in candidates with weak brightness and small scale.Secondly,the region area and centroid of candidates is calculated;Finally,the candidates are matched with the real celestial bodies recorded in the Star-Catalog,and the objects type is recognized based on the category information in the Star-Catalog.