Research On Weakly Supervised Learning Based On Deep Neural Networks In Computer Vision

In the past ten years,deep neural networks have developed rapidly and have made re-markable achievements in various applications.Especially in the field of computer vision,the relevant theories and models for fully supervised learning are optimized continuously,and they have achieved great success in various image recognition tasks.However,with the increasing complexity of various application tasks in the computer vision and the in-creasing amount of data required for various tasks,the labor and material costs required for the data-labeling process are also increasing.At the same time,more real scene datasets are continually emerging,in which the complexity and diversity of images are more sig-nificant,which not only makes it increasingly difficult to thoroughly label related infor-mation,but also the process of labeling is extremely error-prone.It is not very easy to guarantee the accuracy of the label.Therefore,while ensuring the image recognition ef-fect,how to reduce the dependence on data labels and reduce the cost of data labeling has become an urgent problem in the development of deep neural networks.Given the above problems,this thesis will explore weakly supervised learning meth-ods based on deep neural networks in the computer vision field.Compared with fully su-pervised learning methods,it completes the same image recognition task,which requires lower labeling of the data and does not require labels It needs to match the task correctly,and it is simpler in form.The critical issues in this research are(1)how to make full use of the existing information in the data,(2)how to reduce the labeling workload,(3)how to integrate external experience and rules with the model.Specifically,this thesis takes the weakly supervised learning method based on the deep neural network as the primary research goal.By analyzing the advantages and dis-advantages of existing methods,it explores effective weakly supervised learning modeling methods in the field of computer vision applications.Improve the effectiveness of related tasks.This thesis proposes three weakly supervised learning methods based on deep neu-ral networks.The main contents and contributions are summarized as follows:(1)A weakly supervised learning method based on multi-scale evidence is proposed.Multi-scale evidence of the input image is extracted through the pyramid feature hier-archy of the convolutional neural network.Only the global image-level labels are used for weakly supervised learning,which could perform image multi-label classification and point-wise object localization.The method uses a new loss function,which effectively balances the active region on the activation map with different scale categories.A weakly-supervised bounding box generation algorithm based on superpixels is proposed.The ex-perimental results show that the proposed network can efficiently use multi-scale evidence in the image,improve the performance of image multi-label classification and point-wise object localization,and the new loss function can further improve the performance of classification and localization.(2)A weakly supervised learning method based on graph convolutional networks is proposed.It explores how the label dependence in images to help with classification and point-wise object localization.The node vector of the graph convolutional network in the method uses a novel designed initialization method,which removes the dependence of the existing initialization method on the transfer of word embeddings in natural language.It uses the matrix decomposition method to learn the label information in the training set.The experimental results show that the proposed network can effectively use the label dependency in the image to improve the performance of image multi-label classification and point-wise object localization.The proposed vector initialization method can further improve the performance of classification and localization.(3)A weakly supervised learning method based on spatial division is proposed.This method gets rid of the dependence on candidate regions and can achieve end-to-end train-ing.Using only global image-level labels and a novel designed mutual constraint learning process,the model can detect the bounding box directly,which achieves weakly super-vised detection.The original model was transformed into an end-to-end detection network by adding two differentiable modules:Determination Network and Parameterised Spatial Division based on the existing weakly supervised learning network.The bounding box is effectively output explicitly,and the performance is better than the original model on both image multi-label classification and object detection.