Image Semantic Analysis Based On Generative Adversarial Network

Semantic segmentation,as one of the basic and challenging tasks in computer vision,aims at assigning labels to each pixel in an image.For its wide applications in a series of computer vision problems,semantic segmentation has attracted more and more research attention.With the development of hardware such as video card driver,semantic segmentation algorithm based on deep learning has become the mainstream method due to its powerful data fitting ability.However,the limitations of the optimization method based on the segmentation model of deep convolutional neural networks(DCNN)lead to bottlenecks in the final segmentation result.This paper proposes a semantic segmentation method based on generative adversarial network(GAN),and studies it from the following three aspects: 1)We analyze and verify the effectiveness of GAN for semantic segmentation task,and propose a new method called Seg GAN,which uses a trained GAN to help segmentation model get better prediction,since the trained GAN model learned the potential relationship between the label image and the real image.2)This paper analyzes and solves the problem of high-order inconsistency in deep networks.Since the performance of segmentation models based on DCNN is currently affected by high-order inconsistency,it is mainly caused by the inherent spatial transformation invariance of CNN and the independent prediction of pixels.Based on the analysis of this cause,a new hybrid method consisting of segmentation model and GAN model is proposed.This model can solve this problem through long distance semantic continuity.3)There are still deficiencies in the stability of adversarial learning,which will make it difficult for the segmentation model to achieve excellent performance.This paper improves the model's structure,and introduces Wasserstein distance,and solve mode collapse problem.4)We design and implement the image semantic analysis system based on semantic segmentation,which allows users to specify the image to be segmented,and view the visual segmentation result at the result page.Meanwhile,this system allows user download the result file for other semantic analysis task.Extensive experiments on public benchmarking database demonstrate the effectiveness of the proposed method.In summary,this paper designs and implements a semantic segmentation model based on semantic segmentation through in-depth study of the GAN.The model performs the semantic segmentation task of the image with high quality.