Research On Image Semantic Segmentation Based On Convolutional Neural Network

Image semantic segmentation refers to the segmentation of images and the addition of semantic annotations to visually express the information contained in the image region and the relationships between the regions.Therefore,it has an irreplaceable role in scene understanding,analysis,recognition and tracking detection of images.The traditional image segmentation method only performs simple and rough segmentation on the image,and the segmentation result is poor for complex images.Introducing deep learning methods enables end-to-end training of images and pixel-level segmentation and recognition of images.In order to improve the accuracy and accuracy of segmentation,it is of great significance to use the deep learning method to semantically segment images.This paper uses an improved convolutional neural network method to semantically segment images.On the one hand,a SegNet segmentation network is used to semantically segment the image of a specific road scene.The quality of the input image directly affects the SegNet network segmentation result and the single segmentation scenario.In this paper,the input image is preprocessed and the fusion multi-scale method is used to make the network.Ability to learn multi-scale contextual features.On the other hand,the SegNet network model with multi-scale fusion is trained to select the appropriate training set for the multi-scene image semantic segmentation.The network parameters of the model and some functions used in the network,such as optimization function and loss,are constantly adjusted.Functions,etc.The main research work of this paper includes the following three aspects:(1)Segmentation network using SegNet.The quality of the input image directly affects the segmentation result of the SegNet network.In this paper,the input image is pre-processed to denoise the input image,and the useful information of the image is preserved as completely as possible while reducing the noise in the image.The segmentation results before and after preprocessing are analyzed and compared.(2)A multi-scale improved SegNet network model.For the SegNet segmentation network,only the image of a specific road scene can be segmented.For the problem of single scene segmentation,the fusion multi-scale method is adopted to enable the network to learn multi-scale context features,and different sampling rates are used on a given feature layer.The cavity convolution performs efficient resampling,and the results of sampling the respective holes convolution branches are combined to obtain the final result.(3)Training and optimization of the imporved model.In this paper,PASCAL VOC data set is adopted as the appropriate training set of the model,constantly adjusting the network parameters of the model and the functions used in the network.Compare the segmentation results of network models using different functions and select appropriate functions to achieve the effect of optimizing the model.