Research On Object Detection Technology Based On Multi-scale Feature Fusion And Context Analysis

Object detection technology,which identifies and locates objects from a given image or video,is one of the important topics in the field of computer vision,and promotes the development of intelligent surveillance,face recognition and image segmentation.At present,the object detection model based on deep learning has two typical problems: low-precision object detection accuracy and weak adaptive ability in complex environments such as objects overlap occlusion.Therefore,the main purpose of this paper is to analyze the causes of the problem and propose several efficient object detection models,which can effectively solve these two problems and ultimately improve the detection accuracy of the model.Firstly,aiming at the problem of low precision detection of small-scale objects,this paper proposes a object detection model called MS-Faster R-CNN based on multi-scale feature fusion.Based on the process architecture of Faster R-CNN,an optimized multi-scale is adopted.The feature fusion strategy,combined with the FPN pyramid feature,uses two links to complete the feature fusion,making the semantics of the fusion feature more abundant.The cascading RPN and the optimized NMS method are used in the candidate frame recommendation stage,so that the candidate frame of the small-scale obeject is not excessively suppressed,and the recommendation efficiency of the candidate frame is improved.Finally,the ROI Align pooling technique based on bilinear interpolation is used to avoid the loss of precision caused by quantization.Then,aiming at the problem of weak adaptability in complex environment,this paper proposes a context detection model based on context analysis,LSTM-Faster R-CNN,which adopts LSTM-based global context feature extraction method,which can fully express each pixel unit.Global context information.Then,the relationship information between the object relationship component learning candidate frames,that is,the local context information is introduced in the all-connection layer,and the local context information needs to be integrated for the pooled features corresponding to each candidate frame.Object classification and frame detection using features incorporating two kinds of context information can effectively improve the detection accuracy in the object occlusion overlap environment.Finally,in order to verify the validity of the two models,the paper tests on the VOC2007,VOC2012 and MS COCO datasets respectively.The results show that MS-Faster R-CNN can effectively overcome the problem of low-scale object detection accuracy.The results of LSTM-Faster R-CNN detection in complex environments are superior to other models.At the same time,this paper combines two models to propose MS-LSTM R-CNN,and designs and implements a object detection system based on Caffe and Diango.Users can import or collect images after logging in to the system in the browser.By performing object detection on the image,the system can visually display the detection result to the user,and the system proves the practicability of the proposed model.