Research On Object Detection Algorithm For Long Tailed Distribution Problem

With the continuous development of computer vision,deep learning-based object detection techniques have been widely applied in various fields of daily life,such as face recognition,intelligent transportation,and industrial inspection.In scenarios where the number of object categories is small and the distribution among categories is relatively balanced,most object detection algorithms can achieve good detection performance,meeting the detection requirements in these scenarios.However,in natural scenes,objects often follow a long-tailed distribution,where extreme data imbalance severely damages the detection performance of algorithms,resulting in a significant decrease in detection accuracy.The long-tailed distribution problem makes object detection techniques unable to cope with numerous complex natural detection scenarios.By analyzing the characteristics and challenges of object detection algorithms and the long-tailed distribution problem,this paper proposes the following improvements to the twostage object detection algorithm Faster R-CNN from different perspectives:(1)From the perspective of enhancing protection for rare classes,three improvements are made: First,the loss function is reweighted from the perspective of gradients to reduce the excessive suppression gradients exerted by majority classes on rare classes,dynamically balancing the positive and negative sample gradients for each class on the classifier.Second,the non-maximum suppression algorithm is improved from the perspective of instance resampling,using category information and category data quantity as priors to adaptively adjust the intersection over union threshold for different categories,rebalancing the data distribution,and increasing the detection rate of rare class candidate regions.Finally,from the perspective of data augmentation,instance-level and image-level data augmentation methods are combined to increase the number of samples for rare classes,and a category-oriented sampling method is employed to improve the probability of sampling rare classes.(2)From the perspective of better feature representation,two improvements are made by combining attention mechanisms and representation learning: First,the feature extraction network Res Net is improved by using branch attention mechanisms to achieve adaptive adjustment of receptive field size,strengthening the network’s ability to extract and represent useful information,and using deformable convolutions to improve the network’s modeling ability for irregular regions.Second,the feature pyramid network is improved by changing the network structure and integration method between low-level and high-level features,making the output features more balanced,and using self-attention mechanisms to enhance the discriminability of the integrated features.Experiments and ablation studies are conducted on the long-tailed dataset LVIS to compare and evaluate the proposed improvements.The results demonstrate that the proposed methods effectively alleviate the negative impact of long-tailed distribution on object detection algorithms and significantly improve the detection performance of Faster R-CNN in long-tailed distribution scenarios.