The learning of a distance metric, to measure the degree of appearance similarity between a pair of person images, becomes a remarkable challenge in person re-identification due to the unbalanced nature of the datasets, where the number of instances of a query identity is very limited against the vast quantity of different people representations, among which it must be identified. This article presents two network models, a Siamese and a Triplet one, which exploit the multiple possible combinations of training samples in pairs and triplets, respectively. Both models have been used to learn features for nine different body parts, which has been simultaneously analyzed to embed the inter-view variations in a Mahalanobis distance. The influence of the model and training data in the features learning has been evaluated through several tests over the challenging PRID2011 dataset, besides of the proposed system re-identification performance in comparison with other state-of-the-art approaches.

Shape from Focus (SFF) is a passive technique that is used to recover the 3D shape of an object using a series of images with different focus settings. When 2D image sequences are obtained with a specific sampling step size along the optical axis in SFF, mechanical vibrations occur in the position of each image frame. These mechanical vibrations, also referred as jitter noise, affect the accuracy of 3D shape recovery. In this manuscript, the jitter noise and focus curves are modeled as Gaussian function. This is followed by a Bayes filter application, designed to reduce the jitter noise. The filter is applied to each image frame in the Gaussian approximation of the focus curve. The proposed method is experimented by using synthetic and real objects to show performance improvement.

Automated handwritten script recognition is an important task for several applications. In this article, a multi-language handwritten numeral recognition system is proposed using novel structural features. A total of 65 local structural features are extracted and several classifiers are used for testing numeral recognition. Random Forest was found to achieve the best results with an average recognition of 96.73%. The proposed method is tested on six different popular languages, including Arabic Western, Arabic Eastern, Persian, Urdu, Devanagari, and Bangla. In recent studies, single language digits or multiple languages with digits that resemble each other are targeted. In this study, the digits in the languages chosen do not resemble each other. Yet using the novel feature extraction method a high recognition accuracy rate is achieved. Experiments are performed on well-known available datasets of each language. A dataset for Urdu language is also developed in this study and introduced as PMU-UD. Results indicate that the proposed method gives high recognition accuracy as compared to other methods. Low error rates and low confusion rates were also observed using the novel method proposed in this study.

Detection of human beings in a complex background environment is a challenging task in computer vision. Most of the time no single feature algorithm is rich enough to capture all the relevant information available in the image. In this paper, we propose a new feature extraction technique that combines three types of visual information; shape, color, and texture, and is named as the Color space Phase features with Gradient and Texture (CPGT) algorithm. Gradient concept and the phase congruency in color domain are used to localize the shape features. The Center-Symmetric Local Binary Pattern (CSLBP) approach is used to extract the texture information of the image. Fusing of these complementary features yields to capture a broad range of the human appearance details that improves the detection performance. The proposed features are formed by computing the gradient magnitude and CSLBP values for each pixel in the image with respect to its neighborhood in addition to the phase congruency of the three-color channels. Only the maximum phase congruency magnitudes are selected from the corresponding color channels. The histogram of oriented phase and gradients as well as the histogram of CSLBP values for the local regions of the image are determined and concatenated to construct the proposed descriptor. Principal Component Analysis (PCA) is performed to reduce the dimensionality of the resultant features. Several experiments were conducted to evaluate the performance of the proposed descriptor. The experimental results show that the proposed approach yields promising performance and has lower error rates when compared to several state of the art feature extraction methodologies. We observed a miss rate of 2.23% in the INRIA dataset and 2.6% in the NICTA dataset.

In this paper we propose a novel layered-printing method consisting of superposed visible cmy and invisible fluorescent ultraviolet (UV) rgb inks. Our approach can be used to generate a variety of visual color-alteration effects such as revealing two completely distinct images when the print is illuminated with either standard visible or 365 nm ultraviolet (UV) light (Figure 1). This is achieved by computing the maximum achievable color gamuts for both illumination conditions, generating accurate estimates, and applying a spatial-varying gamut mapping to minimize potential ghosting artifacts and calculate the optimal ink surface coverages that, when printed, generate the desired image-alteration effect. Our method uses invisible UV-rgb fluorescent inks which are printed onto a transparent film. It is placed on top of a visible print consisting of standard cmy inks. By separating the UV and the visible inks using the transparent film, physical mixing of the two different ink types is avoided. This significantly increases the intensity of the fluorescent emission resulting in stronger and more vivid color-alteration effects. Besides the revealing of two different images, the same method can be applied for other use cases as well, such as enhancing or adding specific parts to an image under one illumination condition, generating personalized document security features, or aiding color-blind people in color distinction.

A real-time violent behavior detection algorithm based on a new descriptor is proposed. This descriptor reflects a common observation that the changes in both the magnitude and direction of movement in violent images are more abrupt than non-violent ones. During several frames, descriptor feature vectors consisting of descriptor values are generated, and they are inputs to the Support Vector Machine (SVM) classifier for discriminating violent actions from non-violent actions. Comparison experiments among the Motion Binary Pattern (MBP), the Violent Flow (ViF) and the proposed algorithm were conducted with three different types of datasets. In all datasets, the proposed algorithm was above 80% in the F-measure and outperformed the other methods in every case.

In this paper, we propose a new approach to tracking the fingertips of guitarists by embedding a CNN-based segmentation module and a temporal grouping-based ROI-association module combined with a particle filter. First, a CNN architecture is trained to segment hand area of each frame of input video. Then, four fingertip candidates (fore, middle, ring and little fingertips) on each frame are located by counting the vote number of template matching (TM) and reversed Hough transform (RHT). Furthermore, temporal grouping-based ROI association is applied to removal noise and group the fingertip candidates on consecutive frames. Finally, particles are distributed between associated fingertip candidates on every two adjacent frames for tracking the fingertips of guitarists. Experiments using videos containing multiple persons’ guitar plays under different conditions demonstrate that the proposed method outperforms the current state-of-the-art tracking algorithm in terms of the hand area segmentation accuracy (98%) and the fingertip tracking mean error (5.16 pixel: 0.22 cm on the guitar neck) as well as computation efficiency.