Research On Object Detection Methods Based On Weakly Supervised Deep Learning

Object detection is one of the most basic and challenging research topics in the field of computer vision.It has a wide range of applications in robot navigation,intelligent video surveillance,industrial detection and other fields.It also provides the essential support for face recognition and detection,object tracking,pedestrian re-identification and other research fields.In recent years,with the rise of deep convolutional neural network and the great increment of labeled data,object detection based on deep convolution neural network has achieved better performance than traditional methods.Nowadays,the deep convolutional neural network based object detection faces two challenges.The one is that training neural network model requires a large amount of labeled data,which is time-consuming and laborious to obtain and label.And the other is that the datasets of object detection usually have the inter-class imbalance problem which leads to low detection accuracy.Considering the existing problems in object detection,we aim to study the deep convolutional neural network and weakly supervised learning based object detection.And we focus on improving the accuracy of object detection by utilizing the weakly supervised information.The main works of this thesis are listed as follows:(1)Learning object scale with center click supervision for object detection.Recently,the fully supervised object detection methods based on bounding box achieve good detection performance,but they require lots of annotated bounding boxes.Although the weakly supervised object detection methods based on image labels can save the annotation cost,their detection accuracies are relatively low.In view of the existing problems in object detection,the proposed method incorporates CNN visualization with center click annotation to generate the pseudo groundtruths(i.e.,bounding boxes).These pseudo ground-truths can be used to train a fully-supervised detector.On both Pascal VOC 2007 and Pascal VOC 2012 object detection datasets,the proposed method achieves higher detection accuracies.It is noteworthy that the proposed method only use center click annotations as supervision information for achieving a good trade-off between annotation cost and object detection performance.(2)Data augmentation method via weakly supervised semantic segmentation.The object detection datasets always involve the inter-class imbalance problem.Some methods for solving this problem can not achieve high accuracies and may lead to over-fit.Some other methods require the mask annotations for augmenting the datasets.However,annotating the mask is laborious and time-consuming.Considering these problems,the proposed method utilizes the bounding boxes as supervision information for training the weakly supervised semantic segmentation network.And the proposed method uses the masks generated by the trained semantic segmentation network to augment the datasets.The proposed method can reduce the inter-class imbalance rate and increase the diversity of object detection datasets.On the Pascal VOC 2007 and Pascal VOC 2012 object detection datasets,the proposed method achieves higher detection accuracies than the other state-of-art methods.