Weak Supervision Object Location Based On Convolutional Neural Network And Its Application

With the development of hardware and software and the increase of network bandwidth,image and video data are increasing.How to make the computer effectively process and understand these data to assist the relevant industry personnel is an important issue.With the rise of deep learning in 2012,the research focus of industry and academia gradually shifted to the vision algorithm based on convolutional neural network.At present,Convolutional Neural Networks is widely used in all aspects of computer vision.Among them,as a basic research,object localization is applied to many high-level visual tasks,such as intelligent transportation,intelligent security,medical imaging,etc.This thesis focus on "weakly supervised object localization based on convolutional neural networks".In this thesis,a hierarchical feature fusion object localization algorithm is proposed,and a fine-grained image recognition algorithm based on object localization is proposed.The main contents of this thesis are as follows:(1)Based on the ideas of SSD and Dense Net,this thesis proposes a hierarchical feature fusion object localization method.Current object localization methods,such as CAM and SPSM,extract information from the last convolutional layer of the convolutional network.This will lack low-level information.The method proposed in this thesis fuses high-level feature maps and low-level feature maps.It can make up for the lack of object shape and size information in the high-level feature map,and suppress the noise brought by the low-level feature map.The proposed algorithm is compared with CAM and SPSM on the CUB-200 dataset,Caltech101 dataset,and Image Net dataset.On the three data sets,the average localization success rates increased by 3.3%,11%,and 4.7%,respectively,compared to CAM.Compared with SPSM,the average localization success rates increased by 3.7%,1.4%,and 4.7%,respectively.(2)This thesis optimizes the object localization algorithm proposed in this thesis for fine-grained recognition tasks,and proposes a Localization Cutting and Padding Learning framework.The main function of object localization in fine-grained recognition task is to find and enlarge the object in the image.In some images,the objects are very small,which will affect the classification results.The algorithm proposed in this thesis can enlarge these small objects so that the size of all objects is similar.Thus,the recognition accuracy is improved.The relationship between the parts of the object is broken by Cutting and Padding operation,which can make the sub image more independent.But the high-level semantics are not broken.In this thesis,our method is compared with other methods on CUB-200 dataset,Stanford Cars dataset and FGVC Air dataset.The results show that LCPL is superior to other methods.