Deep Learning-based Algorithms For Image Semantic Segmentation

With the continuous development of computer vision technology,image semantic segmentation is widely used in autonomous driving,smart security,medical images,land detection and other fields.Image semantic segmentation refers to assigning its corresponding semantic category label to each pixel in the image,and obtaining the information expressed by the image through the semantic features in the image,which is a key step in understanding computer scenes.Image semantic segmentation based on deep learning can achieve better segmentation effect by proposing the features of each layer in the image.Therefore,the study of semantic segmentation based on deep learning has important theoretical significance and research value.Aiming at the problems of insufficient utilization of shallow features and inaccurate segmentation of small object edges in semantic segmentation,a semantic segmentation network model with iterative feature fusion and MSPS module was designed,and an iterative feature fusion network was built on the Res Net101 basic network to make full use of shallow Layer features and deep features,improve ASPP and SPP modules at the same time,and propose MSPS module to obtain a wider range of context information.For more linear contour objects in the image,a series of morphological operations are used for post-processing.The experimental results show that the performance of the proposed model is significantly improved on the UAV remote sensing dataset UDD6: the OA value reaches 84.14%,and the m Io U reaches74.70%.Aiming at the problems of serious noise interference and redundant information in remote sensing images captured by high-resolution satellites,image enhancement methods such as Gaussian blur and Gaussian noise are firstly used to solve the difficulties of many object categories and low visibility in satellite images.In the above,a remote sensing segmentation network based on cross-stage fusion and channel attention mechanism is proposed.The network pays more attention to the relationship between image channels,while reducing redundant operations makes the network more stable.The experimental results show that the proposed model has an obvious improvement effect compared with the U-Net network.The OA value of the satellite remote sensing dataset GID is increased by 5.82%,and the m Io U is increased by5.37%.