Design And Implementation Of Boundary-Prediction Based Traditional Cultural Image Semantic Segmentation System

Traditional patterns are the crystallization of the wisdom of ancient Chinese people.These patterns carry the aspirations of various ethnic groups for a better life in different periods,which have great inheritance significance and research value.Nowadays,with the rapid development of computer vision technology,how to segment these traditional cultural images in a better way to extract different semantic elements has become an urgent problem to be solved.In this paper,we propose an end-to-end semantic segmentation algorithm based on the current effective deep learning technology to improve the coarse boundary issue of existing segmentation algorithms from two perspectives,and develop a traditional cultural image segmentation system based on this basis.(1)Constructing of traditional culture image dataset.Use image annotation tools to semantically annotate the pre-processed images at the pixel level,save the results and convert their formats.Semantic annotation of 3025 images based on human-computer interaction is completed to form a training set and a test set,which lays the foundation for the subsequent training of semantic segmentation algorithms for traditional cultural images and system practice.(2)To build a semantic segmentation model based on boundary prediction for the problem of coarse boundary of semantic segmentation results.The task of semantic segmentation of traditional cultural images has strict requirements on the quality of the edges of segmented objects.It is also challenging to extract clear boundaries for traditional cultural images due to their diverse textures and complex color spaces.In this paper,an improved iterative upsampling strategy based on boundary prediction is proposed,which can further classify the pixel points by a pre-learned point head at the stage of generating the prediction map.And it can obtain a prediction map with higher boundary quality by combining the low-level features of the image.It is experimentally verified that the improved upsampling prediction module can effectively extract more details and improve the performance of the model.(3)A semantic segmentation model with label relaxation loss is proposed to address the problem of pixel labeling errors.Since it is difficult to achieve pixel-level perfect labeling in the annotation of traditional cultural image data,some improvements are needed to enhance the quality of the results.In this paper,we summarize and analyze the difficulty of boundary labeling when creating datasets,and propose a loss function based on label relaxation for the analysis results,and combine it with the original cross-entropy loss function in the supervised training process of the model.And finally this loss function effectively improves the quality of segmentation results.(4)A semantic segmentation system for traditional cultural images is designed and implemented.The system combines traditional segmentation methods and semantic segmentation algorithms to effectively address a variety of segmentation tasks.The system integrates the Chinese traditional culture material library and classifies the traditional culture images in detail.The system was finally tested in detail to ensure it’s reliable.In this paper,we analyze the current state of the segmentation algorithm and construct related dataset,improve the algorithm from two perspectives of fuzzy boundary and labeling noise,design and implement a semantic segmentation system for traditional cultural images,and satisfy the user’s need to segment and extract traditional cultural elements from images.