Search

Blurred face recognition is still remaining as a challenge task, but with wide applications. Image blur can largely affect recognition performance. The local phase quantization (LPQ) was proposed to extract the blur-invariant texture information. It was used for blurred face recognition and achieved good performance. However, LPQ considers only the phase blur-invariant texture information, which is not sufficient. In addition, LPQ is extracted holistically, which cannot fully explore its discriminative power on local spatial properties. In this paper, we propose a novel method for blurred face recognition. The texture and structure blur-invariant features are extracted and fused to generate a more complete description on blurred image. For texture blur-invariant feature, LPQ is extracted in a densely sampled way and vector of locally aggregated descriptors (VLAD) is employed to enhance its performance. For structure blur-invariant feature, the histogram of oriented gradient (HOG) is used. To further enhance its blur invariance, we improve HOG by eliminating weak gradient magnitude which is more sensitive to image blur than the strong gradient. The improved HOG is then fused with the original HOG by canonical correlation analysis (CCA). At last, we fuse them together by CCA to form the final blur-invariant representation of the face image. The experiments are performed on three face datasets. The results demonstrate that our improvements and our proposition can have a good performance in blurred face recognition.

Video quality assessment (VQA) has been a hot topic due to the rapidly increasing demands in related video applications. The existing state-of-art full reference (FR) VQA metric ViS3 uses adapted the Most Apparent Distortion (MAD) algorithm to capture spatial distortion first, and then quantifies the spatiotemporal distortion by spatiotemporal correlation and a HVS-based model from the spatiotemporal slices (STS) images. In this paper we argue that the STS images can provide enough information for measuring video distortion. Taking advantage of an effective and easy-applied FR image quality model GMSD, we propose to measure video quality by analysing the structural changes between the STS images of the reference videos and their distorted counterparts. This new VQA model is denoted as STS-GMSD. To further investigate the influence spatial dissimilarity, we also combine the frame-by-frame spatial GMSD factor with the STS-GMSD and propose another VQA model, named SSTS-GMSD. Extensive experimental evaluations on two benchmark video quality databases demonstrate that the proposed STS-GMSD outperforms the existing state-of-the-art FR-VQA methods. While STS-GMSD works all square with SSTS-GMSD, which validates that STS images contain enough information for FR-VQA model design.

We propose an automated lung tumor segmentation method for whole body PET images based on a novel downhill region growing (DRG) technique, which regards homogeneous tumor hotspots as 3D monotonically decreasing functions. The method has three major steps: thoracic slice extraction with K-means clustering of the slice features; hotspot segmentation with DRG; and decision tree analysis based hotspot classification. To overcome the common problem of leakage into adjacent hotspots in automated lung tumor segmentation, DRG employs the tumors' SUV monotonicity features. DRG also uses gradient magnitude of tumors' SUV to improve tumor boundary definition. We used 14 PET volumes from patients with primary NSCLC for validation. The thoracic region extraction step achieved good and consistent results for all patients despite marked differences in size and shape of the lungs and the presence of large tumors. The DRG technique was able to avoid the problem of leakage into adjacent hotspots and produced a volumetric overlap fraction of 0.61 ± 0.13 which outperformed four other methods where the overlap fraction varied from 0.40 ± 0.24 to 0.59 ± 0.14. Of the 18 tumors in 14 NSCLC studies, 15 lesions were classified correctly, 2 were false negative and 15 were false positive.

The knowledge of the location of the centers of vertebral bodies and intervertebral discs is valuable for the analysis of the spine. Existing methods for the detection and segmentation of vertebrae in images acquired by computed tomography (CT) and magnetic resonance (MR) imaging are usually applicable only to a specific image modality and require prior knowledge of the location of vertebrae, usually obtained by manual identification or statistical modeling. We propose a completely automated framework for the detection of the centers of vertebral bodies and intervertebral discs in CT and MR images. The image intensity and gradient magnitude profiles are first extracted in each image along the already obtained spinal centerline and therefore contain a repeating pattern representing the vertebral bodies and intervertebral discs. Based on the period of the repeating pattern and by using a function that approximates the shape of the vertebral body, a model of the vertebral body is generated. The centers of vertebral bodies and intervertebral discs are detected by measuring the similarity between the generated model and the extracted profiles. The method was evaluated on 29 CT and 13 MR images of lumbar spine with varying number of vertebrae. The overall mean distance between the obtained and the ground truth centers was 2.8 ± 1.9 mm, and no considerable differences were detected between the results for CT, T1-weighted MR or T2-weighted MR images, or among different vertebrae. The proposed method may therefore be valuable for initializing the techniques for the detection and segmentation of vertebrae.