The computational performance of graphical processing units (GPUs) has improved significantly, achieving even speed-up factors of 10x-50x compared to single-threaded CPU execution are not uncommon. This makes their use for high throughput machine vision very appealing. However, GPU programming is challenging, requiring a significant programming expertise. We present a new programming framework that mitigates the challenges common for GPU programming while maintaining the significant acceleration.

A density-based outlier detection (OD) method is presented by measuring the local outlier factor (LOF) on a projected principal component analysis (PCA) domain from real world spatialtemporal (ST) traffic signals. Its aim is to detect traffic data outliers which are errors in data and traffic anomalies in real situations such as accidents, congestions and low volume. Since the ST traffic signals have a high degree of similarities, they are first projected to two-dimensional (2D) (x,y)-coordinates by the PCA to reduce its dimension as well as to remove noise, while keeping the anomaly information of the signals. Based on the designed LOF algorithm, a semi-supervised approach is employed to label any embedded outliers. It reaches an average detection success rate of 93.5%.

One of the most successful features for texture recognition is the Local Binary Pattern. The LBP is the 8 digit binary number created by comparing the value of a central pixel with its 8 neighbours where 1s and 0s are assigned when respectively the central pixel is larger or smaller than its neighbour. This pattern is bit shifted circularly to its maximum value to obtain rotational invariance. Comparing histograms of LBPs provides leading texture recognition. In our research, we rank the center pixel with all its 8 neighbours. Each pixel is substituted by a 3x3 grid where the numbers one through nine appear once and correspond to the rank of the underlying pixel values (of the local 3x3 neighbourhood) i.e. an input image is transformed to look like a Sudoku grid. Then, we read out the ranks clockwise starting with the right-most rank and appending the central pixel to the end, we then rotate to the maximum value (so achieving rotational invariance). Each 9 digit number is non-linearly mapped to the interval [0,1] so that the overall dataset histogram has a uniform distribution. By comparing the histograms of our Sudoku rank features, we observe a significant increase in recognition performance for the Outex and Curet benchmark datasets.

Traditional image enhancement techniques revise the distribution of pixels or local structure and achieve the impressive performance in image denoising, contrast enhancement and color adjustment. However, they are not effective to improve the overall aesthetic image quality because it may involve contextual modifications, including the removal of disturbing objects, inclusion of appealing visual elements or relocation of the target object. In this paper, we propose a new aesthetic enhancement technique that edits the structural image element guided by a large collection of good exemplars. More specifically, we remove/insert image elements and resize/relocate objects based on good exemplars. Additionally, we remove undesirable regions determined by user interaction and fill these holes seamlessly guided by the exemplars. Based on the experimental evaluation on the database of two landmarks, we observe the considerable improvement in aesthetic quality.

In this paper, we propose a novel hole filling method in view synthesis by using deep convolutional neural networks (DCNN). The hole filling networks are learned by end-to-end mapping between hole regions and ground truth images. Hole regions are initially filled with background information. Subsequently hole filling networks restore high quality of hole filling results. The proposed hole filling networks consist of three layers, which are patch and feature extraction layer, non-linear mapping layer, and restoration layer. Experimental results demonstrate that the proposed DCNN-based hole filling method is able to significantly improve hole filling performance, compared to conventional hole filling methods. Furthermore, responses of filters learned by proposed DCNN show that the proposed hole filling framework could provide visually plausible image structures and textures to hole regions.

Coherent change detection (CCD) images, which are products of combining two synthetic aperture radar (SAR) images taken at different times of the same scene, can reveal subtle surface changes such as those made by tire tracks. These images, however, have low texture and are noisy, making it difficult to automate track finding. Existing techniques either require user cues and can only trace a single track or make use of templates that are difficult to generalize to different types of tracks, such as those made by motorcycles, or vehicles sizes. This paper presents an approach to automatically identify vehicle tracks in CCD images. We identify high-quality track segments and leverage the constrained Delaunay triangulation (CDT) to find completion track segments. We then impose global continuity and track smoothness using a binary random field on the resulting CDT graph to determine edges that belong to real tracks. Experimental results show that our algorithm outperforms existing state-of-the-art techniques in both accuracy and speed.

An emerging poultry meat quality concern is associated with chicken breast fillets having an uncharacteristically hard or rigid feel (called the wooden breast condition). The cause of the wooden breast condition is still largely unknown, and there is no single objective evaluation method or system known for rapidly and non-invasively detecting this quality defect in boneless-skinless chicken breast fillets. Thus, there is an immediate need to develop a rapid and non-invasive sensing technique to detect the wooden breast condition. In this study, sub-surface microstructure and optical properties of poultry meat were measured by optical coherence tomography (OCT) at 930 nm and hyperspectral imaging from 400 to 1,000 nm. The analysis of the measured OCT B scan images showed that the thickness and pattern of the epimysium (the fibrous connective tissue surrounding the muscle tissue) of the meat could be a good feature to differentiate between normal and wooden breast fillets. The OCT signals under the fats and whitish strong connective tissue were smeared with speckle noise so that the epimysium layer edge disappeared under these locations. Because OCT imaging had a small field of view (˜1 cm x 1 cm), it was implied that the scanning time of a large area such as a chicken fillet would be very long. On the other hand, hyperspectral imaging was effective to rapidly scan the entire surface of each fillet and detect excessive fats and strong connective tissue although a spectral analysis showed that there was no pronounced difference between mean spectra of normal and wooden breast fillets. The study results suggested that hyperspectral imaging would increase the throughput of OCT imaging while OCT would detect the wooden breast condition, when both modalities were fused. Thus, a fusion of OCT and hyperspectral imaging will provide a sensing tool to rapidly and accurately detect and sort chicken breast fillets with the wooden breast condition.

Bit-depth expansion is a method to increase the number of bit. It is getting important as the needs of HDR (High Dynamic Range) display or resolution of display have been increased because the level of luminance or expressiveness of color is proportional to the number of bit in the display. In this paper, we present effective bit-depth expansion algorithm for conventional standard 8 bit-depth content to display in high bit-depth device (10 bits). Proposed method shows better result comparing with recently developed methods in quantitative (PSNR) with low complexity.

Establishing dense correspondence fields between images is an important issue with many computer vision and computational photography applications. Although there have been significant advances in estimating dense correspondence fields, it is still difficult to find reliable correspondence fields between a pair of images because of their geometric and photometric variations. In this paper, we propose an unified framework for establishing dense correspondences, consisting of sparse matching, multilevel segmentation, and derivation of affine transformations. Dense correspondence fields are estimated via winner-takes-all (WTA) optimization by utilizing affine transformations, derived from spare matching and multilevel segmentation. The proposed method reduces a size of label search space dramatically, and further extends the dimension of label search space, by leveraging affine transformation with the multilevel segmentation scheme. Our robust dense correspondence estimation is evaluated on extensive experiments, which show that our approach outperforms the state-of-the-art methods both qualitatively and quantitatively.