Research on Optimization Simulation of Image Segmentation Algorithm Based on Generative Adversarial Networks Model

Multi-target tracking and segmentation is to track multiple objects in the video and segment them at the instance level respectively. This task is a novel task integrating classification, detection and segmentation. It not only needs to overcome the overlapping and occlusion of multiple targets in the tracking process to ensure the stability of the tracking trajectory, but also needs to segment each target at the pixel level. With the continuous improvement of accuracy, the standard of video image segmentation is more stringent. In order to solve these problems, this paper carries out research work, and puts forward an optimization model of image segmentation algorithm based on Gan (Generative Adversarial Networks) model to realize automatic segmentation of video images. The improved GAN architecture of multi-scale spatial pooling of this model consists of a generator and a discriminator. Adding AM (Attention mechanism) makes the network pay more attention to the target area and suppress the characteristic response of irrelevant areas. For the discriminator module, CNN (Convolutional Neural Network) structure with multi-scale spatial pooling module is adopted, so that the network can obtain the input of multi-scale fusion features. The simulation results show that the mIOU accuracy of U-Net is only $90.2 \%$, and the mIOU accuracy is improved by $6.6 \%$ after training in the confrontation mode. After adding the method based on GAN in this paper, the segmentation results are obviously improved. The noise irrelevant to the foreground in the segmentation results is obviously reduced, and the purity of the segmented picture is greatly increased, which improves the model segmentation ability under the background interference and clutter conditions. The proposed algorithm improves the robustness of video target segmentation network when it encounters the problems of background interference and clutter of moving targets.