Research On Feature And Class Imbalance Problem In One-Stage Object Detection Algorithm

Object detection is an important research content in the field of computer vision.One-stage detector is of great significance in the field of object detection because of its higher efficiency and lower deployment cost than two-stage detector.In practical application,the main reason that hinders the performance of one-stage object detector is feature imbalance and class imbalance.The feature information in the low and high level of feature pyramid is inconsistent with that in the process of feature fusion,information attenuation and channel attenuation will occur;softmax loss function is good at optimizing the distance between classes,but weak in optimizing the distance within classes,so it can’t solve the problem that the class imbalance affects the detection accuracy well.Based on this,the research of this paper mainly proposes corresponding solutions to these two kinds of imbalanced problems from the perspective of model and algorithm,and puts forward a one-stage object detection framework.The main work and innovations of this paper are as follows:Design progressive feature pyramid network:Feature imbalance in the original feature pyramid network is caused by the attenuation of feature information and the attenuation of advanced feature channels.In view of the attenuation of feature information,the feature polishing module and the fusion module can enhance the original feature representation by fusing the features of high and low levels,and alleviate the problem of the attenuation of feature information.In view of the attenuation problem of advanced feature channels,the number of channels decreases,which reduces the amount of subsequent prediction calculation,but also brings the loss of information.Sub-pixel skipping fusion and sub-pixel context enhancement can alleviate the attenuation problem of feature channels.Therefore,the feature polishing module fusion module is explored Subpixel jump fusion and subpixel context enhancement improve the performance of pyramid network and achieve better performance.Design softmax adaptive loss function:Class imbalance is divided into foreground-background imbalance and foreground-foreground imbalance.Softmax function will have biased gradient estimation,which makes the classification results have errors.Aiming at this problem,Softmax Adaptive Loss(SAL)function is proposed in this paper to solve the class imbalance problem.SAL dynamically adjusts the training process of samples according to the degree of imbalance between positive and negative samples of the classes.SAL can solve the problem of foreground-background imbalance and foregroundforeground imbalance in a one-stage long tail task at the same time.Based on the above two solutions,this paper proposes One-Stage Object Detector(OSOD).In this paper,OSOD is applied to the actual scene,a vehicle detection system is designed,and the practical value of OSOD is verified.