Research On Salient Curves Detection Technology For Cartoon Images

Nowadays,Chinese animation industry flourish.Cartoon images,as one of the basic elements of animation industry,have more and more processing requirements.Image Saliency detection is a basic problem of computer vision and image processing,whose purpose is to identify the main objects in the image which can catch human's visual attention.Curves of cartoon images contain most of the image information.In this background,our dissertation focus on the research of Salient Curves Detection Technology for Cartoon Images,trying to improve detection effects of cartoon images'salient curves by different methods.Firstly,considering the interference of noise to salient curves detection for cartoon images,we propose a salient curves detection algorithm based on texture separation.This algorithm firstly separates textures from structures in a noisy image to obtain a clean structure map.And then detect edges in the structure map using second-order difference algorithm to get the salient curves of the cartoon image.This algorithm greatly reduces the interference of noise to salient curves detection for cartoon images.Then,considering the possibility of losing part of salient curves when the magnitude of textures is near the magnitude of structures in the same cartoon image we propose a salient curves detection algorithm based on structure recovery.This algorithm,taking both distance between curves and length of curves into consideration,filter part of curves which are most likely to be structures from the original cartoon image,and add them to the salient curve map which is incomplete.This algorithm can remove noise and retain relative complete salient curves at the same time for cartoon images whose textures and structures have similar magnitudes.It improves the detection effects.Finally,considering that not all the elements on cartoon images are salient objects and images always follow some composition rule,we propose a salient curves detection algorithm based on composition rules.This algorithm analyzes four composition rules and extracts relative curve features.Then we research the mapping relationships between those features and salience value and use models we trained to detect salient curves for cartoon images which follow some composition rule.This algorithm has good detection effects for cartoon images following some composition rule and can detect curves of real salient objects through all the elements in images.All in all,this dissertation proposes three salient curves detection algorithms and improves detection effects from different perspective.At the same time,the possible developing direction of salient curves detection technology is prospected.