Research On Semi-supervised Method Of Image Semantic Segmentation Based On Segmentation Discriminator Network

Image semantic segmentation is one of the research hotspots in the field of computer vision,Its essence is to classify the pixels in the image,widely used in indoor navigation,indoor robots,driverless car and other fields.At present,the image semantic segmentation of indoor scene method based on deep learning is the mainstream method in the field of computer visual.Due to the feature of the indoor scene,the target occlusion is serious,and the manual access to a rich pixel-level label has gradually become one of the challenging tasks.This essay uses RGB-D image data as the foundation of research.For this reason,we will focus on the above challenges,which contain three research contents:(1)Aiming large amount of pixel-level label data is required on Image semantic segmentation,which will consume a lot of manpower and so on,This essay proposes an image semantic segmentation method based on semi-supervised mode of generative adversarial network.introduced a Image semantic segmentation technology with a loss function based on distance transform and pixel-wise cross-entropy and semi-supervision mechanism.This technology to reduce the use of label data and enhance the accuracy of boundary pixels.(2)Aiming at the noise problem of light intensity and target objects with similar colors and textures in RGB images,This essay proposes a segmentation network for fusing depth geometric information.the features of the RGB image and the depth image are extracted for sparse fusion,which take advantage of improved of the quality of the boundary lines.(3)Aiming at the problem that the pixel holes and noise for the depth,A semi-supervised semantic segmentation of images based on depth completion technology is proposed.This essay proposes a deep completion technology combined with the gated convolution and the SelfAttention mechanism.The technology has deeper extracts depth information,which is effective to supplement the full pixel noise position.In turn,the quality of the image semantic segmentation boundary is improved.Through the comparison of the experimental results on the public datasets NYU-V1 and NYU-V2 with the existing methods.The experimental results show that the proposed method using depth geometric fusion,depth completion technology is verified,thereby improving the problem of segmentation target infection and boundaries unclear.