Image Segmantic Segmentation Based On Superpixel And Mixed Depth Model

Scene understanding can assist or enhance the ability of computers or smart devices to perceive and understand complex and variable scenes.It is widely used in real-world industries such as autonomous driving systems,robot navigation,drones,and wearable devices.Image semantic segmentation is one of the key tasks in scene understanding.It need to segment different objects in the image and assign corresponding semantic labels to each pixel.At present,many image semantic segmentation is based on convolutional neural networks,which are generally based on fully supervised tasks.However,for fully supervised image semantic segmentation,it is very time-consuming and labor-intensive to obtain a large number of pixel-level fine annotations.On the other hand,it is still difficult to search an image semantic segmentation method with high precision and cross-domain and multi-modal data and suitable for any scene.Therefore,the paper aims to find a weakly supervised image semantic segmentation method based on image-level annotation to solve the problem of difficult pixel-level annotation and complex scene understanding.Image segmentation is the basis of image semantic segmentation,and its segmentation quality directly affects the segmentation precision of semantic segmentation.Therefore,the paper focuses on image segmentation and weakly supervised semantic segmentation in image semantic segmentation.The details are as follows:(1)The algorithm that operates using pixels as the basic processing unit in an image segmentation has high complexity,limited extractable visual features,and has a low image segmentation accuracy.The paper proposes an image segmentation method of adaptively generates superpixels.Firstly,the number of initial superpixels is automatically obtained by using the image histogram statistics information,which solve the problem of the initial superpixel number needs to be artificially set.Secondly,the under-segmented superpixels are detected and processed by calculating the maximum difference of the color components in the superpixel,and the visual features are extracted after processing.Finally,on the basis of considering the neighbor relationship between superpixels,the similarity between adjacent superpixels is counted to merge adjacent superpixels with the highest similarity.And the superpixel clustering center is updated by averaging.(2)For the problem of how to effectively map image-level labels to pixels in an image under the weakly supervised semantic segmentation conditions.The paper proposes a weakly supervised semantic segmentation algorithm based on candidate region and depth mixed model.The algorithm first obtains target candidate regions in the image based on superpixel segmentation algorithm.Then,different neighborhood particles are generated by the neighborhood rough set for the candidate region.According to image-level labels statistical information,the candidate region semantic labels are inferred from the highest frequency semantic label,and other are inferred based on the strongest semantic association relationship.Finally,the extreme learning machine(ELM)is trained using the candidate region with semantic labels to obtain accurate and fast classification result,which can reduce the introduction of negative sample pixels in the training data.In addition,the ELM does not require iterative adjustment parameters and can be added to new datasets for training at any time,which can improve algorithm efficiency and generalization ability.