Semantic Segmentation Algorithm Of Image Based On Non-Local Points Relations

As an important information carrier,images play an important role in both daily life and scientific research.Image segmentation is a branch in the field of artificial intelligence.In machine vision technology,it is an important part of image understanding.The essence of semantic segmentation is to solve the problem of image segmentation,which aims to divide each pixel into differ-ent areas according to artificially defined semantic categories,thereby dividing areas with human's interest.As a low-level vision task,the semantic segmen-tation task has been widely used in scene understanding,autonomous driving,computer-aided diagnosis,image editing and other fields.With the development of deep learning,semantic segmentation algorithms based on convolutional neural networks have became a popular research di-rection.The structural characteristics of convolutional neural networks lead to insufficient long-term dependency of contextual information,which affects the quality of segmentation.Calculating the relationship of non-local points can obtain more long-term dependent information and significantly improve the segmentation performance.Therefore,this paper establishes the non-local relationships between points on the high-dimension features of the middle layer in the neural network and the final segmentation result,which have been applied to the actual skin area detection task of portrait.The main innovative research results are as follows:(1)A semantic segmentation algorithm is proposed,which is based on non-local relation model in high-dimension feature space.The algorithm first uses the dual distance information of high-dimension features to efficiently model non-local relationships in a feature map.Secondly,a spatial attention module is proposed to enhance object-level feature consistency.At the same time,an auxiliary loss function is proposed to enhance the category-level feature consis-tency.Through the consistency enhancement from two aspects,the continuity of high-dimension features is ensured,and the continuity of the segmentation results is also improved.The proposed method surpasses other relation module on three datasets,while significantly reducing the amount of calculation.(2)A semantic segmentation post-processing optimization algorithm is proposed based on non-local relations in the segmentation results.Firstly,a di-rected graph model suitable for the calculation in neural networks is established on the segmentation results to represent non-local relationships,Secondly,an efficient message transfer mechanism is established on the model to optimize the segmentation results.Finally,the solution process of the graph model is de-signed as a general semantic segmentation post-processing module,which can be inserted into the existing semantic segmentation model to perform the end-to-end prediction.It can improve the segmentation result in low-confidence regions.For efficient training,a fast training strategy for the proposed post-processing module is also proposed.(3)For the task of skin color classification and beautification,the proposed semantic segmentation algorithm is used to detect the skin area.First,in order to find the problems in current methods,the existing datasets and algorithms are used to train the model.Secondly,optimization plans for the current model and data are formulated to improve the semantic segmentation model proposed above united with downstream applications Then a multi-supervision and multi-data joint training strategy is proposed to obtain accurate and edge-fitting skin regions,which can support the skin color correction and grading tasks.