Picture Segmentation Algorithm Based On Local Region Conditional Random Field Model

Semantic segmentation of images[1-2,3]technology is the cornerstone technology of image understanding.In fact,the so-called semantic segmentation of images technology is used to separate different objects in the image into different region blocks,in order to classify each pixel in the picture.The research achievements have been widely applied in various directions,such as autonomous driving technology,medical imaging technology analysis,remote sensing image analysis and other fields.The semantic segmentation of images technology has been developing slowly due to technical reasons in the past decades.However,as the research boom of deep learning,convolutional neural networks have gradually become the core method of computer vision assignment research.It is contributed that the powerful capability of feature extraction.In addition,it is evitable that carrying out the features extraction and reconstruction for original image data.The full convolutional neural network proposed is derived from the development of convolutional neural networks.In this method,convolutional neural network structure pioneered has been applied in semantic segmentation assignments,transcending the traditional image segmentation algorithm in segmentation accuracy and segmentation speed.In general,the major work of semantic segmentation task is to extract the position information of image space and the dependence between pixels and pixels.This paper conducts in-depth research based on convolutional neural network from specific application scenarios.A novel semantic segmentation model structure is proposed.This paper can be divided into four parts as follow:1.In-depth study of traditional semantic segmentation techniques is presented,including semantic segmentation results based on full convolutional neural networks and semantic segmentation structures based on conditional random field models.Then the advantages and disadvantages of the two semantic segmentation models are compared and analyzed.2.A new DeepLab-Res18 model based on full convolutional neural network is proposed.This model combines the latest convolution structure-ResNet structure,which to some extent avoids the problem of losing spatial position information for the object due to pooling effect in the original convolution process.Moreover,the cavity convolution technology has been used in the model,in order to raise the receptive field size effectively with the guarantee for image resolution.3.The application scenarios of the model are exactly analyzed.A semantic segmentation network called local area conditional random field model based on conditional random field model is proposed,which fully considers the adaptability between the application scenario and the model.The novel algorithm is used to optimize the iterative process of the conditional random field model,so that the model can extract the dependence between the pixel and the pixel in the picture,adding a smooth process to the segmentation result and optimizing the segmentation result at the boundary region of the object.4.In order to effectively utilize the strong correlation information between consecutive picture frames and frames,this paper designs a convolutional neural network tracking algorithm,which can obtain the information between adjacent frames of pictures and use this information pair.The segmentation result of the adjacent frame picture is error corrected.The experimental results show that the local area conditional random field model described in this paper can be applied to the semantic segmentation of traffic scene pictures in a timely and high-precision manner.