The Research Of Human-machine Face Cognitive Consistency In Interactive Face Retrieval

With the development of the information society as well as science and technology, the technology of face image recognition and retrieval has been widely used in security, video monitoring, criminal cases, etc. In practical applications, there exists semantic gap between low-level features and high-level semantic cognition, so, the interactive technology is introduced to narrow the gap and speed up the retrieval efficiency. Retrieval system involves two key techniques: image feature and distance measure. While the distance of the similarity measure is directly related to the image features. So, the study of features seems particularly important. We explore the face cognitive consistency between human and machine mainly from the aspect of face feature, by using the theory of human cognitive methods for feature selection, extraction and reconstruction. The main research contents are as follows:1. The problem of human-machine face cognitive consistency in interactive face retrieval is expounded. Experiments on large number of feedback data collected from users in interactive face retrieval system are performed to analyze the consistency probability distribution. The probability distribution shows a non-uniform decreasing trend, indicates that human-machine face cognition has some degree of consistency. In addition, we analyze the human-machine face cognition consistency under different number of display images in system, results show that too little or too much display images will reduce the performance of the system.2. A new feature selection method SRFS(Sparse Representation for Feature Selection) is proposed, SRFS maps a linear combination of the feature space to category value. By utilizing truncated Newton interior-point method to obtain nonnegative sparse solution, SRFS use the sparse values to select class-distinguished feature subset, such as identity, gender and race relevant sub-features from face images. By exploring the influence of the three sub-features on human-machine face cognitive consistency, we obtained the face similarity retrieval more inclined to identity. By comparing the three sub-features to the original features characteristics, it shows that the face similarity cognition of human considers many aspects.3. Manifold learning method is introduced in the interactive face retrieval system. We propose two new feature dimension reduction algorithms SR-LLE(Sparse Representation-the Local Linear Embedding) and SR-LE(Sparse Representation –Laplacian Eigenmaps), which fuse Sparse Representation with manifold learning method. Experiments on face identity, gender and race recognition verify the effectiveness of the proposed algorithm. On the study of consistency, we further proposed the a method to reconstruct features based on LLE(Local Linear Embedding- Chong Gou, LLE- CG), the experimental results show that the manifold features can do well in recognition task but it does not necessarily can improve human-machine cognitive consistency.