As the most common type of facial occlusion, eyeglasses may cause great accuracy degradation in face recognition. In this paper, we proposed an improved approach on automatically detecting and removing eyeglasses from grayscale facial images. First, we normalized a face image by the result of face detection and eyes localization. Then we used a fast judging method to detect eyeglasses’ presence. For images with eyeglasses, we used PCA reconstruction error and edge feature to determine the occluded area, and synthesized the area through image inpainting. Experimental results show that our approach can detect the presence of eyeglasses very accurately and obtain generally natural looking images without eyeglasses. In face recognition test, our approach greatly contributed to the accuracy of recognition, achieving higher improvement than other approaches such as simple PCA reconstruction, iterative error compensation, and weighted fusion.

In practical problems, there are usually no clear counterparts as reference to evaluate restoration results. So no-reference blur assessment is very important and necessary. In this paper, we proposed an objective measure named as Edge Factor (EF) to appraise image blurring. The fundamental rationale was that blurring effect was much more perceptible in edge transition zones. The pixel number of edge transition zones would decrease when blurring occured. We defined the pixel number ratio of the edge transition zones to the whole image as EF. Experimental results show the monotonic consistency of EF and RMS. The proposed method is further compared with some common edge detection algorithms to demonstrate the effectiveness of combining point-based entropy with Pulse Coupled Neural Network.

Anaglyph 3D is a low-cost and popular 3D display method, made out of a photo pair. With only one common camera, however, there are difficulties in capturing moving objects. We suggest using the light field camera, which can be exploited as a computational stereo rig, to capture image pairs simultaneously. A processing framework for making anaglyph 3D image is then proposed. The effectiveness of our proposal is verified visually by experiments.

Face detection, pose estimation and facial landmark localization are three fundamental problems in pattern recognition. These three tasks have high request of algorithm efficiency and accuracy. Zhu and Ramanan proposed a model based on mixture of tree structures to solve the three tasks simultaneously and it obtains state-of-the-art result. However, the efficiency of their algorithm is relatively low. Our improved algorithm combines Viola Jones detector and tree-structured model and achieves a speed-up of tens of times even hundreds of times of original algorithm on ordinary laptop according to images of different sizes.