Synthesized Heterogeneous Face Sketches Recognition And Assessment

With growing attentions paid to information security, face recognition technology has more and more applications. Sketch-based recognition is an important branch of face recognition. Sketch-based face recognition relies on the performance heavily of face sketch-photo hallucination. Here face sketch-photo hallucination means transforming sketches and photos to a same modality, Viz. transforming a sketch to a photo or vice versa. However, this transforming process may result in different extents of distortions. Then, it is important to perform face recognition based on synthesized images according to image quality between different image pairs. In addition, since there are many different face sketch-photo hallucination algorithms, it is difficult to determine which algorithm performs better in terms of image quality. Therefore, image quality assessment to synthesized images is the key issue.Firstly, this thesis proposed a face recognition method on synthesized images through image quality assessment(IQA). This method filters an image by difference-of-Gaussian filter and then IQA methods are explored to evaluate the image quality to obtain the quality scores for recognition. The proposed is training-free and has a low time complexity. Experiments on multiple databases illustrate the effectiveness of this method.Secondly, this paper constructs a synthesized face sketch database for IQA. There are only face sketch-photo pairs in existing face sketch databases without synthesized sketches. Moreover existing IQA databases are mostly for natural image with different distortion types and different degrees of distortions but not for face sketches which has structure property and other special characteristics. The constructed database contains face sketches and synthesized sketches by five different face sketch hallucination methods on two databases: CUHK Student database and AR database. Moreover volunteers are invited to subjectively evaluate the quality of these synthesized sketches.Finally, considering the different structure characteristics between face sketches and natural images, IQA methods for natural images cannot be directly applied tosynthesized sketches. And then an IQA method for synthesized images based on feature similarity was designed in this thesis. The gradient features and edge features from Canny edge extractor are deployed and they are fused to calculate IQA scores. Experimental results demonstrate that it achieve promising results on synthesized sketches.