Research On Scene-based Graphic Element Extraction And Application

The massive increase of image data in the current era has promoted the rapid development of machine vision technology.It is of great significance to use technical means to organize and analyze image data to obtain effective image information.The information of an image is mainly composed of various image sub-elements,such as tables and chairs,people,animals,etc.in ordinary family scenes,various main respects in Thangka images,etc.These image sub-elements are picture elements.Different scene images have different elements,but the underlying features of the elements are similar.This paper proposes a set of schemes based on the current deep learning target detection technology,which is suitable for effectively extracting scene primitives in different scene images.The main contents of this article include:(1)An improved model based on the SSD algorithm is proposed,which improves the accuracy of the model.After combed and analyzed the current development process and technology direction of the current target detection technology,summarized the characteristics and results of each technology,proposed a primitive element extraction model based on the SSD algorithm,and optimized the basic network and loss function to make the model accurate.The mAP value of the native SSD algorithm in the PASCAL_VOC2012 data set increased from 72.4% to 76.1%.The head network is customized using the parameterized scheme,which makes the scheme easy to modify in the application,balancing the task size and the number of layers of the head network.(2)Through innovative data augmentation methods and training optimization strategies,the efficiency and generalization ability of the scheme are effectively improved.The picture data preprocessing and a series of data enhancement methods of this solution are explained,and the innovative method of using data heat supply replaces the traditional way of expanding data first and then inputting it to the model,further improving the effect of data enhancement..During the model training phase,the setting of the learning rate and the changes in the training cycle were optimized to increase the training efficiency of the model,and a mixed-precision training scheme was adopted,which reduced the memory usage by 37% in the model training.(3)This solution is applied and verified the Thangka scene,and the cultural background of the Thangka image is introduced.From the preparation of the data set to the realization of transfer learning,the application method of this solution in different scenarios is explained.The experimental results and some details are analyzed,and the generalization performance of this scheme is proved.