Design And Implementation Of Object Detection Model On TensorFlow

The technology of object detection is a very important subject in computer vision task.It has been widely used in intelligent transportation,face detection,aerospace and medical image equipment.However,in the practical application process,people will find that there are still many factors that will limit the actual detection effect,including the size of the object to be detected,deformation,occlusion,background region transformation and so on.Therefore,in the field of computer vision,obj ect detection is still a scientific research subject that needs to be improved.In this paper,two kinds of object detection algorithms based on regression and region proposals are studied.In object detection based on regression represented by YOLO network,we find that this kind of detection algorithm is fast and simple,but its detection accuracy is not perfect enough.As the representative of region proposals object detection algorithm,Faster R-CNN has advantages in detection accuracy,but it needs a relatively large time cost.We use a popular deep learning framework—TensorFlow,as the experimental platform.We propose a multi-scale target detection algorithm based on RPN region proposal network,which can extract the features of large and small objects by using the level of feature map.In addition,this paper also improves the classification regression model and proposes a two-dimensional loss function,which makes the region proposals closer to the groundtruth boxes in the final training.This paper redesigns and optimizes some of the convolution layers based on YOLO.We choose RPN instead of YOLO's final fully connection layer,and introduce the concept of anchor boxes to regress the location information of the object to be detected,which makes the training process of the improved network easier.The experimental data set used in this paper is PASCAL VOC data set,the accuracy and speed of each category in 20 detection objects are calculated and analyzed.A number of experiments have proved that the proposed multi-scale object detection algorithm based on RPN and the improved YOLO algorithm for small object detection have improved in accuracy and speed.The experiments also verify the effectiveness and feasibility of the two object detection algorithms.