A fast object tracking approach based on sparse representation

Study of school accidents reinforces need for coordination and basic motion training: School Sports Accidents: Analysis of Causes, Modes, and Frequencies Kelm J, Ahlhelm F, Pape D, et al. J Pediatr Orthop. 2001;21:165–168. Study conducted at Chirurgische Abteilung, Krankenhaus Neunkirchen/Saar, Akademisches Lehrkrankenhaus der Universitdt des Saarlandes, Germany

This paper presents a new approach to single-image super-resolution, based on sparse signal representation. Research on image statistics suggests that image patches can be well-represented as a sparse linear combination of elements from an appropriately chosen over-complete dictionary. Inspired by this observation, we seek a sparse representation for each patch of the low-resolution input, and then use the coefficients of this representation to generate the high-resolution output. Theoretical results from compressed sensing suggest that under mild conditions, the sparse representation can be correctly recovered from the downsampled signals. By jointly training two dictionaries for the low- and high-resolution image patches, we can enforce the similarity of sparse representations between the low resolution and high resolution image patch pair with respect to their own dictionaries. Therefore, the sparse representation of a low resolution image patch can be applied with the high resolution image patch dictionary to generate a high resolution image patch. The learned dictionary pair is a more compact representation of the patch pairs, compared to previous approaches, which simply sample a large amount of image patch pairs, reducing the computational cost substantially. The effectiveness of such a sparsity prior is demonstrated for both general image super-resolution and the special case of face hallucination. In both cases, our algorithm generates high-resolution images that are competitive or even superior in quality to images produced by other similar SR methods. In addition, the local sparse modeling of our approach is naturally robust to noise, and therefore the proposed algorithm can handle super-resolution with noisy inputs in a more unified framework.

According to the image processing theory, complex texture image regions have larger entropy values, an adaptive steganography algorithm in the sparse domain is proposed here. The algorithm first divides an image into nonoverlapping image blocks, then computes the entropy of each block and chooses the mean of all entropies as a threshold. Image blocks whose entropy is higher than the threshold are selected for sparse decomposition as complex texture image regions. The secret message is embedded into the decomposition coefficients, and then the stego image is reconstructed with modified coefficients. The experimental results prove that, given the same payload, the proposed adaptive algorithm has a better performance against the steganalyzer.

To acquire the better object tracking performance in this paper, the object of interest is modeled by its RGB-color histogram feature together with deep convolutional neural network feature (deep CNN feature) and sparse representation. Because the information of the objects of interest cannot be obtained ahead in practical use, the deep CNN features should be abstracted through the pre-trained VGG network. Then the deep feature and RGB-color histogram are combined to model the object in adaptive mode. And then sparse representation is used to express the model. Above all, an adaptive particle filter algorithm based on deep CNN feature together with RGB-color feature and sparse representation is proposed to track the object of interest. Extensive experiment results demonstrate the effectiveness of our proposed method under the serious object occlusions and appearance changes.

We propose a fast kernel sparse representation based classification (SRC) for undersampling problem, i.e., each class has very few training samples, in face recognition. The proposed algorithm exploits a nonlinear mapping to map the data from the original input space into a high-dimensional feature space. Then, it performs very fast sparse representation and classification of samples in this space. Similar to the typical SRC methods, the proposed approach is based on the L1 norm minimization, whose direct solution can be very time-consuming. In order to improve the computational efficiency, our method uses the coordinate descent method in the feature space, which can avoid directly solving the L1 norm minimization problem, and significantly expedites the computational procedure. Compared with other SRC methods based on the L1 norm minimization, our proposed method achieves very high computational efficiency, without significantly degrading the classification performance. Several experiments on popular face databases demonstrate that our method is a promising efficient kernel SRC based method.

Tensor-based sparse representations of multi-phase medical images for classification of focal liver lesions

Multi-task l0 gradient minimization for visual tracking

Deep unfolding methods design deep neural networks as learned variations of optimization algorithms through the unrolling of their iterations. These networks have been shown to achieve faster convergence and higher accuracy than the original optimization methods. In this line of research, this paper presents novel interpretable deep recurrent neural networks (RNNs), designed by the unfolding of iterative algorithms that solve the task of sequential signal reconstruction (in particular, video reconstruction). The proposed networks are designed by accounting that video frames' patches have a sparse representation and the temporal difference between consecutive representations is also sparse. Specifically, we design an interpretable deep RNN (coined reweighted-RNN) by unrolling the iterations of a proximal method that solves a reweighted version of the l1 - l1 minimization problem. Due to the underlying minimization model, our reweighted-RNN has a different thresholding function (alias, different activation function) for each hidden unit in each layer. In this way, it has higher network expressivity than existing deep unfolding RNN models. We also present the derivative l1 - l1 -RNN model, which is obtained by unfolding a proximal method for the l1 - l1 minimization problem. We apply the proposed interpretable RNNs to the task of video frame reconstruction from low-dimensional measurements, that is, sequential video frame reconstruction. The experimental results on various datasets demonstrate that the proposed deep RNNs outperform various RNN models.

Image representation is a fundamental issue in signal processing, pattern recognition, and computer vision. An efficient image representation can lead to the development of effective algorithms for the interpretation of images. Since Marr proposed the fundamental principle of the primary sketch concept of a scene, many image representations have been developed based on this concept. The primary sketch refers to the edges, lines, regions, and others in an image. These are characteristic features, which can be extracted from an image by transforming an image from the pixel-level to a higher level representation for image understanding. Many techniques developed for pattern recognition can be used for image transform and resulted in efficiently representing the characteristic features (or called intrinsic dimension). Dimensionality reduction (DR) and sparse representation (SR) are two representative schemes frequently used in the transform for reducing the dimension of a dataset. These transformations include principle component analysis (PCA), singular value decomposition (SVD), non-negative matrix factorization (NMF), and sparse coding (SC). The study of the mammal brain suggests that only a small number of active neurons encode sensory information at any given point. This finding has led to the rapid development of sparse coding which refers to a small number of nonzero entries in a feature vector. Hence, it is important to represent the sparsity for the dataset by eliminating the data redundancy for applications. The ultimate goal is to have a compact, efficient, and compressed representation of the input data.

Sparse coding based visual tracking: Review and experimental comparison

Presented is a deep-learning method for pattern classification and object recognition. The proposed methodology is based on an optimised version of the hierarchical temporal memory (HTM) algorithm and it preserves its basic structure, along with a tree structure of connected nodes. The tree structured scheme is inspired by the human neocortex, which provides great capabilities for recognition and categorisation. The proposed method is enriched with more representative quantisation centres using an adaptive neural gas algorithm, and a more accurate and dense grouping by applying a graph clustering technique. Sparse representation using L1 norm minimisation is embedded as a liaison between the quantisation centres and their grouping, reinforcing the proposed technique with advantages, such as a natural discrimination capability. The proposed work is experimentally compared with the aforementioned techniques as well as with state-of-the-art algorithms, presenting a better classification performance.

This paper describes a novel system for real-time video texture analysis. The system utilizes hardware to extract second-order statistical features from video frames. These features are based on the Gray Level Co-occurrence Matrix (GLCM) and describe the textural content of the video frames. They can be used in a variety of video analysis and pattern recognition applications, such as remote sensing, industrial and medical. The hardware is implemented on a Virtex-XCV2000E-6 FPGA programmed in VHDL. It is based on an architecture that exploits the symmetry and the sparseness of the GLCM and calculates the features using integer and fixed point arithmetic. Moreover, it integrates an efficient algorithm for fast and accurate logarithm approximation, required in feature calculations. The software handles the video frame transfers from/to the hardware and executes only complementary floating point operations. The performance of the proposed system was experimentally evaluated using standard test video clips. The system was implemented and tested and its performance reached 133 and 532 fps for the analysis of CIF and QCIF video frames respectively. Compared to the state of the art GLCM feature extraction systems, the proposed system provides more efficient use of the memory bandwidth and the FPGA resources, in addition to higher processing throughput, that results in real time operation. Furthermore, its fundamental units can be used in any hardware application that requires sparse matrix representation or accurate and efficient logarithm estimation.

The paper is devoted to the search for new approaches to encrypting selected objects in an image. Videos were analyzed, which were divided into frames, and in each video frame, the necessary objects were detected for further encryption. Images of objects with a designated geometric shape and color characteristics of pixels were considered. To select objects, a method was used based on the calculation of average values, the analysis of which made it possible to determine the convergence with the established image. Dividing the selected field into subregions with different shapes solves the problem of finding objects of the same type with different scales. In addition, the paper considers the detection of moving objects. The detection of moving objects is carried out based on determining the frame difference in pixel codes in the form of a rectangular shape. Cellular automata technology was used for encryption. The best results were shown by the transition rules of elementary cellular automata, such as: 90, 105, 150, and XOR function. The use of cellular automata technologies made it possible to use one key sequence to encrypt objects on all video frames of the video. Encryption results are different for the same objects located in different places of the same video frame and different video frames of the video sequence. The video frame image is divided into bit layers, the number of which is determined by the length of the code of each pixel. Each bit layer is encrypted with the same evolution, which is formed by one initial key bit sequence. For each video frame, a different part of the evolution is used, as well as for each detected object in the image. This approach gives different results for any objects that have a different location both on the video frame image and in different video frames. The described methods allow you to automate the process of detecting objects on video and encrypting them.

Compact or efficient representation for either images or image sequences is key operation to performing image and video processing tasks, such as compression, analysis, etc. The efficiency of an approximation is evaluated by the sparsity measure of the approximation, i.e., the sparsest the representation is, the more efficient it is. For image processing tasks, it is often desired to decompose the image into a linear combination of few visual primitives or features selected from a large collection of waveforms, called the dictionary. These primitives are usually designed in such a way to achieve some good approximation performances by assuming a given class of functions to model images. A common class of functions for image modeling is the set of functions having discontinuities along smooth contours or boundaries, delineating smooth geometrical regions. It was shown that the well-known separable isotropically scaled two-dimensional wavelets fail to capture the geometrical regularity inherent to images. Motivated by this issue, we designed two rich dictionaries Ds and Dst, composed of multi-scaled ridge-like functions satisfying the anisotropy and the directionality features, for image and video expansions, respectively. Regarding the spatio-temporal dictionary Dst, we defined a warping operator W, which aligns these functions along the coherent motion trajectories in order to exploit the nature of the temporal evolution in the video signal. To obtain a compact representation for the visual signal over the dictionaries Ds and Dst, only the primitives that best match the signal components must be selected. This problem is well studied in approximation theory and it is known as the problem of sparse approximation in unrestricted dictionaries, whose optimal solution is NP-hard. Greedy approaches, such as MP and OMP, have been proposed to provide sub-optimal solutions by iteratively choosing one atom at a time. However, the computational complexity of these algorithms is cumbersome and limits their applicability. To alleviate this issue, we introduced a greedy algorithm, called the M-Term Pursuit MTP, whose performances are very close to those related to MP. Then we employed both algorithms, MP and MTP, for image and image sequence decompositions and we investigated their performances. A target application has been studied, which is the scalable compression, where the aim is to generate a bit-stream that can provide both SNR and geometrical scalability. Geometrical scalability refers to the spatial scalability in case the of images and to the spatio-temporal scalability in the case of video sequences. The SNR or quality scalability is related mainly to the nature of the greedy approaches and the embedded quantization and coding. To do so, we designed a rate allocation algorithm that offers a progressively refinable bit-stream, based on the subsets approach. On the other hand, the geometrical scalability is fullfiled thanks to the structure of the dictionaries, Ds and Dst, which are designed using parametric fashion. Comparisons with state-of-the-art scalable and non-scalable codecs illustrate the good performance of the proposed compression techniques at low rates.

In this paper we propose a new motion compensated prediction technique that enables successful predictive encoding during fades, blended scenes, temporally decorrelated noise, and many other temporal evolutions which force predictors used in traditional hybrid video coders to fail. We model reference frame blocks to be used in motion compensated prediction as consisting of two superimposed parts: one part that is relevant for prediction and another part that is not relevant. By performing prediction in a domain where the video frames are spatially sparse, our work allows the automatic isolation of the prediction-relevant parts. These are then used to enable better prediction than would be possible otherwise. Our sparsity induced prediction algorithm (SIP) generates successful predictors by exploiting the non-convex structure of the sets that natural images and video frames lie in. Correctly determining this non-convexity through sparse representations allows better performance in hybrid video codecs equipped with the proposed work