Application Research On Weakly-supervised Learning In Computer Vision

With the application of deep learning in computer vision,the current deep neural network models perform well in tasks such as image classification,object detection,and semantic segmentation.However,training of the supervised-learning models requires a large amount of manual label.Manual labels,especially location-related labels,often consume a lot of manpower and resources.Therefore,weakly-supervised learning meth-ods,which relies less on labels,have become a research hotspot.Weakly-supervised learning is a way of machine learning.Different from the supervised learning models,whose labels requires one-to-one correspondence with output of the model,the anno-tation that weakly-supervised learning relies only needs information on partial levels.Therefore,weakly-supervised learning has application prospects and practical signifi-cance in industrial computer vision tasks.This thesis bases on the view above to carry out the application research of weakly-supervised learning in computer vision.Training multi-class network model by multi-instance learning is a common idea to construct weakly-supervised model,which could complete classification task while ren-der heatmaps represent the locations of objects,so that it can finish vision tasks based on such heatmaps.However,the object features in the realistic scene images are diverse.How to construct the model to extract the various features is the key content of this the-sis.In addition,the thesis implements the specific computer vision tasks based on the proposed model.The main content of this thesis is as follows:(1)Utilizing the characteristics of different depth features of deep convolutional net-work to extract the semantic information,a network model based on multi-scale feature map is proposed.The corresponding multi-scale global pooling method and method of generating category heatmap are proposed.Besides,this the-sis introduces a constrained loss function makes the heatmap more compact on the object.This model extracts the diversed features through the perspective of multiple semantic levels.(2)By investigating the variants of convolution,combined with the deformable con-volution network,and using different characteristics from different dilation ra-tios for convolution sampling receptive fields,a network model based on multi-dilation-rate deformable convolution is proposed.This model uses class activa-tion mapping and proposed fusion method to generate category heatmaps.This model extracts diversed features through the perspective of multi receptive fields.(3)The proposed models are trained by multi-instance learning method,and ap-plied to multi-category classification,object localization and object detection.To complete object detection task,a detection framework based on the category heatmap is proposed.This thesis compares the performance of proposed models to the related ones,besides the advantages and shortcomings of the model are evaluated by visual analysis.The training on proposed models only relies on image-level labels,and the thesis finally conducts experiments some location-based tasks such as object localization and object detection.The results of experiments prove that the proposed model and method are effective and weakly-supervised learning in compter vision can be applied.