Study On Similarity Calculation And Quality Assessment For Sketch Images

With the popularity of painting education software on smart devices, developing artificial technology to evaluate the quality of painting is important for attracting more users and improving the self-education procedure. A key problem of quality assessment for sketch images is how to extract suitable feature and build a rating model. In the feature extraction stage, we need to solve the problem caused by uncertain sketch location and various size of line weight, so the shape characteristics of images is able to capture. And in the rating stage, we should build a model that can comply with or near human judgment. Unfortunately, the existing models can not fulfil the two requirements. Therefore, this paper focuses on in-depth research on quality assessment of sketch images, we have designed a method to extract multi-scale HOG feature from the images, and quality assessment algorithm for sketch images based on rankingSVM. Generally, there are two major contributions in our work.First, we have designed a method to extract multi-scale HOG feature from the images, which can capture the overall shape of the image effectively. Currently, in most feature extract approach, an image may be described by a set of detected local key points, called bag of word features. However, those local descriptors capture more on local texture, not the shape directly, so the key feature may not be found by using bow model. To solve this problem, the visual perception should be considering when design a feature algorithm, so we are going to design a method to extract multi-scale HOG feature from the images by using the spatial pyramid, which is able to capture the shape characteristics of images from the coarse-grained and fine-grained level. The experiment result show that, multi-scale HOG feature outperform single-level HOG feature, and the more level pyramid feature have, the more shape feature can capture.Second, in the learning stage, by implementing weighting constraints in the model, we are able to improve the learning ability of rankingSVM. Many approaches formulate the training problem as a regression problem or a multi-class classification problem. However, these approach can overlook some essentially useful information, such as the labels, which are dependent or inherent. In this article, we improve the learning ability of rankingSVM by change the objective function of rankingSVM, which also can reduce the difference between ranking values of the pictures that have same score. And finally, we proposed a scoring strategy using a Guassian model based on distribution of ranking values. The result shows that our approach outperform the regression or the multi-class approach.