Research And Implementation Of Efficient Det Target Detection System Based On Bidirectional Feature Fusion

Object detection is one of the most fundamental issues in computer vision,and has received worldwide attention and research,2001 to 2012 is the era of traditional machine learning,representative algorithms include VJ detector,HOG detector,DPM model.In 2012,deep learning was born,leading the object detection to a new height,With the rapid development of deep learning,target detection models emerge in endlessly,the current object detection framework is roughly divided into two categories-single-stage target detector and two-stage object detector,of which the earliest representative of two-stage object detector is R-CNN in 2014,and the earliest representative of single-stage object detector is YOLO at the end of 2015,Then,the research of target detection is based on these two ideas and continuously optimizes and derivates to improve the performance of the model,among which the two-stage target detector has good performance in accuracy.At the same time,feature extraction backbone networks are also emerging one after another,among which VGG,Res Net and so on are representative.This paper selects the two-stage object detector proposed by the Google team in 2020,based on the Efiicient Det of Bi FPN feature fusion network as the baseline,and focuses on reconstructing the feature fusion network,changing the backbone feature extraction network for experiments,and building a target detection system based on the separation of front and back ends of B/S architecture,the main contents are as follows:1.Improve the feature fusion structure.The BiFPN structure is derived from improving PANet and has strong performance.Based on existing feature fusion structures.In this thesis,three feature fusion methods are used for experimentation.2.Change the feature extraction network.In order to ensure the rigor of the experiment,the Res Net backbone feature extraction network was selected for control experiments to verify the effectiveness of the reconstructed feature extraction network.3.Development of an object detection system.Vue and Spring Boot technology are used to build a target detection system based on B/S architecture separation of front and back ends,and realize the effect of object detection by web interaction.