Images Semantic Segmentation Based On Deep Convolutional Neural Networks

Image semantic segmentation is one of the most important algorithms in the field of image processing and computer vision,and is the key technology to assist the computer to recognize the content of the image.The result of semantic segmentation plays a fundamental role in the following tasks,such as scene analysis,target location and tracking.Therefore,it is of great practical significance to study an effective image semantic segmentation algorithm.With the development of deep learning,the high accuracy brought by neural network makes it widely researched and applied in many scenes,such as image recognition and target detection.Compared with the traditional method of semantic segmentation based on region feature extraction,the feature extracted from deep convolutional neural network has better expression,so the algorithm has higher accuracy.The basic idea of semantic segmentation based on deep convolutional neural network is to extract every pixel's semantic features by neural network,and then classify and recognize each pixel by these features,so as to get the semantic segmentation map of the whole image.Therefore,the key of this kind of methods is the accuracy rate of the pixel recognition.This paper is from this perspective,through the analysis of existing semantic segmentation methods based on deep convolutional neural network,combined with the technology of machine learning and image processing,an improved algorithm is proposed,and the effectiveness of the algorithm is proved by experiment.First of all,in order to improve the feature extraction ability of deep convolutional neural network,a multi-scale feature fusion method is adopted in this paper.The full connection layer of the network obtains the semantic features which classify the objects,and the convolution layer gets the details of the object.Combining the two features can enhance the expression intensity of the feature.Therefore,this paper designs a multi-scale pooling network structure,using multiple pool templates to obtain the multi-scale features of objects under different receptive fields,enhance the recognition ability of the network to the target,and improve the accuracy of semantic segmentation.Besides,because of sampling operations in deep convolutional neural network,the structure and location information of the target are lost,resulting in reducing theclassification accuracy of the edge pixel.In this paper,boundary points reclassification algorithm is proposed.Based on the prediction of convolutional neural network,combined with the method of region division,the pixels with low prediction credibility are reclassified,which further improves the effect of semantic segmentation.Finally,the image semantic segmentation method based on deep convolutional neural network is applied to target detection and scene recognition,which shows the application value of the algorithm.In the target detection task,the semantic segmentation method is used to get the location and category attributes of the target,then locate and identify the target,so as to achieve the detection.In the scene recognition task,the semantic segmentation method is used to capture the basic properties of the object's structure,location,category and so on,so as to provide technical supports for scene analysis and target detection.