Research On Imbalance Problems In Deep Object Detection

Object detection is one of the basic tasks in the field of computer vision,which is simultaneously judging the category and location information of objects in a given picture.It has a wide range of applications in surveillance,autonomous driving,medical decision making,and many problems in robotics.Thanks to the rapid progress and development of deep learning technology in recent years,more and more people have devoted themselves to the research of object detection algorithms,and the object detection has also ushered in another high.Compared with model architectures,the training process has received relatively less attention in object detection.In this paper,we carefully revisit the standard training practice of detectors,and find that the detection performance is often limited by the imbalance during training process,which generally consists in three levels:the feature imbalance int the feature pyramid,the imbalance of positive and negative samples in the region proposal networks and the gradient contribution imbalance of easy and hard samples in the loss function.Then we proposed corresponding improvement measures for these three types of imbalances from the perspective of models and algorithms.The main work is as follow:Aiming at the problem of feature imbalance in the feature pyramid,that is the imbalance of imbalance of detailed information and semantic information between high and low feature layers,a residual-group balanced feature pyramid network is proposed.For the problem of feature information loss,we propose residual feature enhancement module.By fusing multi-layer convolution features and performing residual connection feature enhancement operations on high-level features,the feature layer information loss problem is compensated.For the existing features on the problem of semantic gap between feature layers,a grouped feature enhancement module is proposed to generate features with balanced feature information and strong robustness through group feature fusion and spatial attention mechanism.Finally,experiments prove that the improved algorithm can solve the feature imbalance problem in the feature pyramid network,and significantly improve the performance of the existing detection algorithm.Aiming at the imbalance of positive and negative samples in the regional proposal network,a multi-level hierarchical regional proposal network is proposed.Multi-level stratified regional sampling is carried out based on the intersection of the difficult and easy samples and the real marked box compared to the interval distribution,which balances the ratio of positive and negative samples in the regional proposal network,making the sampling distribution for negative samples more reasonable,and the selected samples are more representative in view of the imbalance problem of difficult and easy samples in the smooth L1 loss function,a guided multi-task loss function and a shrinking regression loss function are proposed to balance different tasks and samples in the training process,and improve the gradient contribution of easy samples in the training process.Speeds up the model convergence.The experimental results show that the improved algorithm can effectively solve the problem of imbalance between positive and negative samples and the imbalance of difficult and easy samples in the regional suggestion network and loss function,effectively improving the detection ability of the target,and solving the problem of target missed detection,misdetection and inaccurate positioning frame to a certain degree.