Research On Algorithm Of Image Semantic Understanding Based On CRF

Image semantic understanding is now an extremely important and ultimate goal of the foundation in the field of computer vision.The research results have been widely applied in robot navigation,unmanned,image retrieval,virtual reality,and people's livelihood.It highlights the important practical value and academic value.Each branch task of image-semantic understanding,such as target detection,image segmentation,image classification and so on,understanding the target overall and split the target area outline,but the overall understanding of image semantic is still under study.The current mainstream method commonly used probabilistic graphical models to improve pixel calibration tasks,it is the accuracy of image understanding.The above methods need to solve two problems: firstly,how to extract efficient features,thereby improving the accuracy of pixel classification;secondly,how to combine the results of the pixel preliminary classification with probabilistic graphical models,use constraints graph model to reduce misclassification to improve the accuracy of the final result of image understanding.In response to these two issues,this paper proposes multi-scale image deep learning for image understanding,at the same time,combine deep learning with conditions random field for image understanding.Details are as follows:Firstly,for efficient feature extraction and image pixel classification problem,this paper presented a new multi-scale image semantic understanding deep learning method.The research method used convolution neural network to obtain images of dense features,it is very robustness for the scale changes of image,compared with current methods.In order to improve the accuracy of pixel classification,this paper presents a combination of deep learning and super-pixel segmentation method,improve pixel classification accuracy and depict the outline of the target area,get a good classification results,it show that the accuracy of the method.Secondly,the smoothness constraint of condition random field model is introduced to deep learning framework.The paper presented the loss function of convolution networks under the smooth constraint.The smoothness constraint of pixel classification is added to the iterative process of convolution neural network to improve the speed and efficiency of deep learning.It can effectively solve the problem of inaccurate target boundary contour.