Research On Automatic Segmentation Based On Bi-level Model In Traditional Costume Image

Chinese traditional costume is known as the quintessence of Chinese culture and the representative of Chinese costume.And it is the precious wealth created by the Chinese nation and even the human society.Extracting the representative pattern gene in the traditional costume image is helpful to explore its cultural connotation and discover its cultural inheritance mechanism.This paper takes traditional costume image as the research object,carries out the research on the automatic segmentation algorithm of traditional costume image,and extracts the valuable pattern gene.Although interactive segmentation algorithm has been widely concerned by researchers for its flexible interactive method,high precision and less user interaction work and deep learning has also achieved outstanding results in the field of image segmentation,in the absence of image data sets labeled pixel by pixel,it is still an urgent problem to realize the automatic segmentation of traditional costume image and ensure a high segmentation accuracy.Compared with the natural image,the pattern genes in the traditional costume image are mostly small-size objects with rich styles.Even the pattern genes in the same category are quite different from each other,and on the cloth of different material,the texture of same pattern gene also can have difference.These characteristics increase the difficulty of the segmentation of traditional costume image.A weakly supervised image segmentation algorithm based on a bi-level model is proposed to realize the automatic segmentation of traditional costume image in this paper.The bi-level model is composed of two layers,namely,the object detection layer and the segmentation layer.The model detects the pattern gene obj ect to be segmented before segmentation,which can improve the segmentation precision of small object.Moreover,if the object detector has good generalization performance,it is likely that the detector can detect the same category of pattern genes with large style differences.In the traditional costume image segmentation task,compared with the method of directly training the image segmentation model,the segmentation method based on detection may have better segmentation effect.The object detection layer of the proposed model employs the improved GRP-DSOD++detection framework in this paper to obtain the position coordinates and category information of objects.The segmentation layer obtains the class-agrnostic segmentation results based on the location information output by the object detection layer,and the results corresponding to the category information to obtain the segmentation results with semantics.The main innovations and research contents of this paper include:(1)This paper analyzes the limitations of the existing automatic image segmentation algorithms in the application of traditional costume image,in order to accomplish the task of automatic segmentation for traditional costume image,a bi-level model image segmentation algorithm based on deep learning and interactive segmentation is proposed.The algorithm first outputs the object location and category information through the object detection method based on deep learning,and then inputs the information into the interactive segmentation method to accomplish the automatic segmentation of traditional costume image,and the results have semantic information.(2)Improved GRP-DSOD object detection framework based on deep learning.First,dilated-Inception module(DI module)proposed in this paper is applied to the object detection system,and SE block is introduced to enhance the channel characteristics which are of great use to the back layer and suppress the useless channel features.Second,weight factor is introduced to solve the problem of positive and negative sample balance in the confidence loss function of the loss function.Third,standard volumes in the last two dense blocks of the original network structure are also introduced.At last,the improved object detector is called GRP-DSOD++object detection framework.(3)The image segmentation system based on the bi-level model is designed and implemented.The effectiveness of the system is verified on the traditional costume image data set,and the simulation experiments are carried out on the public data set.