Fast Synthesis Methods And Style Classification Algorithms For Face Sketches

Face sketch synthesis is the process of synthesizing face sketches from face photos by machine learning.It is widely used in the practical applications including law enforcement and digital entertainment.Take law enforcement as an example,it can assist police to narrow down and arrest potential suspects by the automatic search of their photos.In the absence of the frontal and clear photos of suspects,the artists draw the sketches of suspects according to the descriptions of victims or eyewitnesses or the unclear images of surveillance videos.Under the circumstances,the face sketches can be synthesized and recognized in light of face sketch synthesis.Since there are large number of face sketches in police station,and these sketches are likely to be drawn by different artists with different styles,it needs to identify the style of the test sketch firstly.Gallery mug shot photos are transformed into sketches using the training sketches that share the same style with the query sketch.However,existing face sketch synthesis methods are not efficient enough,and to the best of our knowledge,there is no research in the automatic classification of facial sketch styles recently.They cannot satisfy the above demands.Aiming to research in the face sketch synthesis algorithm in a fast manner and the classification of facial sketch styles,the major contributions of this dissertation are summarized as follows:1.A face sketch synthesis algorithm based on dual-transfer is proposed.Existing face sketch synthesis methods assume that the structures of face sketches are same as that of face photos in low-dimensional manifold.And the structure information of the test photo and the training photos is learned and mapped from photo space to sketch space.Then the synthesized face sketch is obtained.But the above mentioned assumption is too strong.When it comes to the identity-specific information in the test data,not in the training data,it is hard to synthesize this information.Considering this fact,we propose a dual-transfer face sketch synthesis algorithm.The first step is the inter-domain transfer.The knowledge of the relationships between the test photo and the training photos is learned,and then transferred from the photo domain to the sketch domain.The second step is the intra-domain transfer.The knowledge of relationships between the training photo-sketch pairs is exploited,and then transferred to the test.The last step is to develop an efficient and effective synthesis process.We formulate the dual-transfer framework as low-rank optimization problem,which has a closed-form solution.Experiments demonstrate that the proposed method can synthesize face sketches with more identity-specific information in a fast manner,compared with the current state-of-the-art.2.A face sketch synthesis algorithm based on compositional model is proposed.Existing face sketch synthesis methods divide a face into patches,and integrate synthesized sketch patches by exploiting weighted average operator.It restrains edges and high-frequency information to some degree,and it is hard to generate the clear texture on the overlapping region.Considering this fact,we simulate the observation of artists before drawing,and propose a compositional model-based face sketch synthesis algorithm.The first step is the template generation.It decomposes a face into seven components rather than divides it into patches in existing methods.And the integrity of the component is guaranteed.The second step is the component generation.We extract the features from the components and feed them into classifier.Thus,the proposed method is divided into the training and test phases.The training phase is performed off-line.It can effectively improve the algorithm on speed.The last step is the component warping and blending.The multilevel B-spline method can fine tune the shape of the components.And it exploits Poisson blending method to fuse the fine-tuned components and avoids edge inhibition problem generated by weighted average operator which is exploited by the existing methods.Experiments validate that compared with the state-of-the-art,the proposed method can not only run faster,but also produce synthesized sketches with more clear and vivid textures.The further experiments on three applications in digital entertainment demonstrate the effectiveness of the proposed algorithm.3.A classification of facial sketch styles based on handwritten components and features is proposed.Existing methods of the classification of styles focus on Chinese paintings or Western oil paintings,which are different from facial sketches.They cannot be applied to solve the problem of the classification of facial sketch styles.Considering above fact,we design a classification of facial sketch styles based on handwritten components and features.The first step is to design handwritten components.In light of the way how art critics to recognize the face sketch artists from components,we design five handwritten components.The second step is to design handwritten features.The artists often employ the different techniques to build their own unique style.In the line with it,we design four handwritten features.The third step is the selective ensemble method based on the support vector machine acting as the classification of facial sketch styles.It enhances the generative ability of the machine learning system and improves the accuracy of the classification.Experiments demonstrate that the proposed algorithm can effectively and efficiently classifies the facial sketch styles.4.A classification of facial sketch styles based on multiple kernel learning is proposed.Since the different features have the different structures with heterogeneous information,the distribution of the data fed into the classification is uneven in the high-dimensional space.Considering above fact,we present a classification of facial sketch styles based on multiple kernel learning.The different kernel functions map the different features into the different feature space.And the heterogeneous data in the combination of the different feature space obtain a better representation.The accuracy of the classification is improved.Compared with the cross-validation,the feature selection and the selective ensemble,the proposed method has some advantages.Experiments validate that the proposed method can not only classifies the facial sketch style effectively,but also synthesizes the different styles of facial sketches with the help of the face sketch synthesis.