Object Detection Based On Improved Faster R-CNN

In the field of computer vision,target detection is the basis of many research directions.The task of the target detection algorithm is to detect the object class in the current scene and to locate the object.Target detection algorithms involve many disciplines such as image processing,artificial intelligence,machine learning,and pattern recognition.Moreover,the target detection algorithm is the basis of many algorithms of computer vision such as face detection algorithm,vehicle detection algorithm,pedestrian detection algorithm and the like.In practical applications,it also covers many fields,such as video surveillance,autonomous driving,industrial,and agricultural medicine.Therefore,the research on the accuracy and real-time of the target detection algorithm has quite a wide range of research significance and practical application value.At present,the mainstream target detection algorithms are mainly divided into two categories: one is the target detection algorithm based on various regional nomination algorithms.The algorithm usually first extracts the pre-selected regions by various methods,and then targets the targets based on the pre-selected regions.Identify and locate.At present,the target detection algorithms based on regional nomination in the mainstream target detection algorithms mainly include: R-CNN,Fast R-CNN,Faster R-CNN,and so on.Another type of target detection algorithm is an algorithm based on regression target detection.This type of algorithm usually uses the target detection problem as a regression problem.The algorithm detects the target as a bounding box and a class probability end-to-end based on the deep learning algorithm.In the above two types of target detection algorithms,the target detection algorithms based on the regional nomination algorithm are generally lower indetection speed than the end-to-end target detection algorithm.However,in the accuracy of target detection,the target detection algorithm based on the regional nomination algorithm usually has better results.In the Faster R-CNN algorithm,a new regional nomination method is proposed instead of the traditional regional search method.The classical convolutional neural network ZF or VGG is used to extract image features and the sliding window is used to generate features.In this paper,we use the new feature extraction network ResNet network and FPN network,which utilizes the multi-layer features of the image,thereby improving the target detection rate,especially the detection rate of small targets.In the Faster R-CNN algorithm,multiple nominated areas are predicted for the sliding window position using an anchor box.The design of the anchor box is based on manual design.In this paper,we improve the anchor box and use the clustering method to automatically design the anchor box.In this paper,we use GIOU as the optimization criterion and loss function of the target regression box to replace the traditional IOU and loss function.Through these improvements,we have achieved the goal of improving the detection rate of small targets.