The Research On Small Object Detection Method Based On RetinaNet

As a research hotspot in object detection,small object detection has a wide range of applications in remote sensing detection,product defect detection,medical image detection and other fields.In recent years,there have been many studies on small object detection at home and abroad,but small objects still have problems such as weak feature extraction capabilities,high requirements for prediction and positioning accuracy,low signal-to-noise ratio of contextual semantic information,and redundant shallow feature calculations.The object detection algorithm of the thesis does not perform well on small objects.Aiming at the above problems,this thesis uses the single-stage object detection model Retina Net as the basis,and improves the four aspects of feature extraction fusion,regression prediction,data enhancement and detection head to improve the detection performance of the method for small objects.The specific contents are as follows:(1)Aiming at the problems of weak feature extraction ability of small objects,easy loss of feature fusion semantics and high requirements for prediction and positioning accuracy,an attention-based two-way path feature fusion small object detection model—BD-Retina Net is proposed.First,the SA-Net attention mechanism module is introduced into the residual module of the Retina Net backbone network to improve the model’s ability to acquire semantic information of small objects.Secondly,a self-adjusting weight multi-scale two-way fusion feature pyramid structure DB-FPN is proposed to increase the weight ratio of small object features in the feature fusion process.Finally,optimize the regression prediction loss function,balance the number of positive and negative samples of the small object prediction frame,and improve the positioning accuracy of the prediction frame.(2)Aiming at the problems of low definition of small object details,low signal-to-noise ratio of contextual semantic information,and redundant calculation of shallow features,a small object detection model based on slice assistance and cascaded sparse query—SAQ-Retina Net is proposed.First,introduce the idea of slicing in image segmentation,use slice-assisted finetuning to increase the proportion of small objects in the image,and use slice-assisted inference to enhance the performance of small object detail features;secondly,design a backbone network AS-Res Net with better performance,using The Switchable Atrous Convolution structure and the Poly-Scale Convolution structure can be switched to reduce the pollution of background noise on small object context information;finally,a detection head based on cascaded sparse query is designed to use the rough position information of small objects in deep features to guide shallow feature generation Sparse feature map reduces redundant calculation of the background area by the model during shallow feature processing.Finally,this thesis uses three public data sets to verify the feasibility and effectiveness of the above two models.The comparative experimental results show that the two detection methods proposed are superior to mainstream detection methods,and can effectively improve the small object detection accuracy of the model.