Research On Key Technologies Of Scene Understanding Based On Machine Learning

Scene understanding is a very active research direction in machine learning and pattern recognition.As the main research content of image scene understanding,although the research work of image classification and image semantic segmentation has made breakthrough progress with the continuous development of machine learning methods,there are still some problems need to be solved.Taking image classification as the main research object and machine learning method as the main research means,this dissertation mainly focuses on the recognition of low classification rate image categories,the time-consuming problem of visual dictionary learning,the multi-task integration design problem of scene understanding,and the spatial complexity and stability of image classification networks in deep learning model.The main innovations are as follows:1.An image classification method based on principal node analysis is proposed to solve the problem of low classification rate image categories by optimizing the image feature expression model.By clustering and filtering image features,this method can greatly reduce the synonymy problem between visual words,and effectively improve the low classification rate of some image categories caused by unbalanced distribution of image categories,intra-class differences and inter-class blurring.2.An image classification method based on fast dictionary learning strategy is proposed to solve the problem of visual dictionary learning efficiency in image classification task.This method optimizes the visual dictionary generation module of image classification model,adopts parallel computing mode,introduces similarity function to sparse sample feature descriptor set,and realizes fast learning of visual dictionary by clustering algorithm based on non negative matrix factorization.This method can greatly reduce the learning time of visual dictionary,significantly improve the efficiency of dictionary learning,and effectively improve the recognition rate of image classification.3.An image classification algorithm based on parallel kernel principal component analysis network and an image classification method based on attention pyramid residual network are proposed respectively to solve the design problem of image classification network based on deep learning model.The former focuses on the spatial complexity of image classification network model.By building a parallel kernel principal component analysis network,it can effectively improve the recognition rate of image classification while solving the problem of high spatial complexity of the classical convolutional neural network model.The latter improves the classical residual network and introduces attention mechanism and spatial pyramid model into residual module,which can effectively improve the stability of deep learning network.The experimental results on the public image classification datasets and the internal datasets of the author's work unit verify the effectiveness of the two proposed network models.4.A scene understanding model based on dual-channel and multi-feature analysis is proposed to solve the problem of multi-task integrated design.By fully mining the various information contained in the color image and its corresponding gray-scale image,the color channel model and gray-scale channel model are established,the multi-source features are extracted by integrating the semantic information in the foreground and background,and the feature dimension of multi-source fusion features is reduced by using the compressed sensing method to realize the dual tasks of image classification and image segmentation.All the proposed methods have been verified by simulation experiments on image classification and image semantic segmentation related data sets and self built image data sets.The experimental results show that the proposed methods can effectively improve the performance of scene understanding subtasks such as image scene classification and image semantic segmentation.