Research And Application Of Small Object Detection Based On Faster R-CNN

Object detection can detect specific targets in images by combining target localization and classification.As one of the important tasks of computer vision,it is widely used in security,traffic and medical fields,such as face detection,autonomous driving and medical image analysis.Until now,the object detection technologies based on deep learning have outperformed human beings in many visual applications,making deep learning algorithm become the research focus of object detection.Small target refers to the target whose resolution is less than 32 ×32 pixels in the image.Due to its few pixels and low recognition accuracy,small object detection has become a research difficulty in object detection studies.Compared with large targets,small targets of images carry less information and bear higher costs in labeling objects.However,conventional object detection algorithm can not deal with the problems of small object detection well.Focusing on multi-scale detection,positive and negative samples balance,and contextual information,this thesis studies small object detection under different application scenarios based on Faster R-CNN.The main results are as follows:(1)Since target sizes are small and features are difficult to extract,this thesis proposes a Multi-scale Spatial Feature Enhancement Pyramid(MSFEP)algorithm.MSFEP deletes the max pooling layer of FPN to enhance the spatially-rich shallow layer features,and improves the detection performance of small target by fusing with semantic-rich deep layer features.The multi-scale anchor generated on MSFEP algorithm and Multi-scale Anchor Generation(MAG)mechanism is more suitable for the shape and size of the target to be detected,so that the Faster R-CNN can detect targets at different scales.Besides,the weight decay regularization method is introduced to reduce the network weight and improve the object detection performance of the detection model.To verify the validity of the proposed algorithm,a face mask detection dataset is constructed upon the face images and face mask images from two public datasets,MAFA and FDDB.Results of face mask detection show that the improved Faster R-CNN based on MSFEP performs better and improves the small object detection performance by 6.4% compared with the original one.(2)A data augmentation strategy based on positive and negative sample balance is proposed to solve the problem that small targets have fewer positive samples in the candidate region generation phase.Small targets carry less valid pixels in the image dataset and provide less information for the model to update the gradient values,which is the direct reason for poor performance in small object detection.The data augmentation strategy selectively copies small targets in the image and pastes them into the training image multiple times to generate an image dataset for small object detection.In the training phase of the model,a mosaic data augmentation algorithm is introduced to enrich the detection background of small targets and enhance the robustness of the Faster R-CNN.Introducing self-attention mechanism,setting the feature extraction network of Faster R-CNN to Swin-Transformer Tiny(Swin-T)network to improve the detection performance.Traffic sign detection results show that the data augmentation strategy based on positive and negative sample balance can effectively improve the small object detection performance of Faster R-CNN.(3)An invasive diagnostic model for bladder cancer is designed based on Faster R-CNN combining MSFEP algorithm with Swin-T network.To solve the problem of less information available for small targets,a data labeling method based on contextual information is proposed to construct a small object dataset for bladder cancer subtypes detection.Then on T2-weighted MR images,MSFEP algorithm and Swin-T network is used to enhance the spatially-rich features between bladder wall and tumor,and border regression loss function with Intersection over Union(Io U)information is used to improve the accuracy of lesion localization task.The class activation mapping visualized by the Smooth Grad-CAM++method not only provides clinicians with an interpretable basis for invasive diagnosis of bladder cancer,but it also indicates that the proposed algorithm can learn the subtypes information of bladder cancer.The experimental results show that the computer-aided diagnosis model can provide effective auxiliary information for clinical diagnosis of bladder cancer invasion.Based on Faster R-CNN,this thesis resolves a series of key issues in small object detection,in order to expand the application scope of object detection.The research results are helpful for improving the performance of small object detection,broadening the application range of object detection,and providing reference for the research and application of small object detection based on Faster R-CNN.