Semantic Segmentation Of Urban Environmental Images Based On GAN

With the development of computer hardware and the rapid rise of artificial intelligence,neural network has become one of the key technologies in various intelligent related solutions.It can be seen that neural network has made a series of remarkable achievements in many aspects,such as image processing,natural language processing.It can extract abstract and high-level features from low-level raw data features by simulating animal brain structure.These advanced features are the portrayal of the nature of the data.Through the processing of these advanced features,great achievements can be made in many realities such as face recognition,smart driving and voice translation.In recent years,the traditional automobile industry has started to enter the field of automatic driving technology.By means of analysising the semantic information of the urban scene,neural network can be applied in collision warning for pedestrians and vehicles,these can help make the control strategy construction for vehicle system.In 2014,Long et al,at UC Berkeley proposed a fully convolutional neural network,replacing the full linked network with the convolution network,achieved a shift from image classification to dense pixel prediction.The effect is not very satisfactory,because the pool layer will lose part of the location information while aggregating the last convolution result.For image semantic segmentation,the location or context information of each pixel is indispensable,which is very important for the final classification of pixels.As a result,later researchers have proposed three different structures to solve this problem: hole convolutional structure,encoder-decoder structure,spatial pyramid structure.This article analyzes the working principles and charact-eristics of several different structures,we compared the differences between various networks.And then,we merge the GAN structure to image semantic segmentation.The innovation points of this paper mainly include the following points:1.Introduce GAN into the field of image segmentation.Combining the idea of conditional GAN,using the original image as the input of the generator,the generator generates the required semantic segmentation results using the original image as the input of the generator.The original images and the semantic segmentation results generated by the generator or the original images and the artificial segmentation results are combined as the input of the discriminator.Training the network until the discriminator can not distinguish the pictures generated by the generator from the human annotations.Then the generator can generate a satisfactory image segmentation result.2.Merge the super pixel information to inputs.In this article,the boundary information obtained by super pixel segmentation is as input of the generator network.We first use SLIC's super pixel segmentation method to obtain the subtle contours of the input image,and then the super pixel segmentation results and the original image are concated together to segmentation result.3.Reconstruction of the new encoder-decoder structure.We found that the segmentation results in the Pix2 pix model can not be clearly classified at the boundaries,so we changed the generator's output layer to classify several channels to output the results.We adopted the encoder-decoder structure and changed the way of pooling indexes to concat features from different level.