Balancing Positive And Negative Samples For Weakly Supervised Object Detection

Object detection can provide classification and location information for image understanding.It has great application value and is widely used in autonomous driving,remote sensing image detection,and other tasks.With the development of deep learning and neural network in recent years,significant progress has been made in object detection methods.The development of fully supervised object detection methods relies on large-scale datasets with precise annotation,but obtaining box-level labels is time-consuming.Therefore,weakly supervised object detection methods have gradually attracted attention.Weakly supervised object detection methods that only require image-level labels greatly reduce the labeling cost of the training datasets.This thesis proposes weakly supervised object detection methods with image-level labels.Weakly supervised object detection has the following problems: due to the lack of box-level labels,it is impossible to divide proposals into positive or negative samples according to the Io U with box-level ground truth like fully supervised object detection methods and select positive and negative samples according to a preset ratio to solve the imbalance;when the image contains multiple objects of the same category,only the top-scoring proposal is selected as the initial positive sample,and the pseudo labels of other objects are marked as background,which is not conducive to the training of subsequent branches;the network tends to focus on the most discriminative region of the object.First,this thesis suppresses the influence of negative samples by hard suppression and soft suppression according to context score.The context score of the proposal is obtained from the weakly supervised semantic segmentation results.Besides,the box regression branch is combined into weakly supervised object detector to simplify the training process.Second,this thesis proposes the top-scoring proposal accumulation strategy based on the inconsistency of the network prediction results at different training stages.As the training progresses,more initial positive samples of the same category in the image can be mined to further balance the number of positive and negative samples.By joining this strategy,reliable pseudo labels can be generated to promote instance classifier refinement branches and box regression branch and improve the detection ability of the network.Third,this thesis further optimizes the feature extraction network with the inconsistency of the different convolution modules' output feature maps.The region of interest in the feature map is closely related to the selected top-scoring proposals.Due to the optimization,the feature extraction network can focus on more complete parts of the objects and make the objects more outstanding,so as to further promote the detection ability of the network.This thesis conducts experiments on two public datasets.Experimental results demonstrate that the methods proposed in each chapter are effective and can improve the detection ability of the network.Compared with other state-of-the-art methods,our method can achieve comparable performance.