Dimche Kostadinov

Fast content identification in high-dimensional feature spaces using Sparse Ternary Codes

We consider the problem of fast content identification in high-dimensional feature spaces where a sub-linear search complexity is required. By formulating the problem as sparse approximation of projected coefficients, a closed-form solution can be found which we approximate as a ternary representation. Hence, as opposed to dense binary codes, a framework of Sparse Ternary Codes (STC) is proposed resulting in sparse, but robust representation and sub-linear complexity of search. The proposed method is compared with the Locality Sensitive Hashing (LSH) and the memory vectors on several large-scale synthetic and public image databases, showing its superiority.

On accuracy, robustness, and security of bag-of-word search systems

In this paper, we present a statistical framework for the analysis of the performance of Bag-of-Words (BOW) systems. The paper aims at establishing a better understanding of the impact of different elements of BOW systems such as the robustness of descriptors, accuracy of assignment, descriptor compression and pooling and finally decision making. We also study the impact of geometrical information on the BOW system performance and compare the results with different pooling strategies. The proposed framework can also be of interest for a security and privacy analysis of BOW systems. The experimental results on real images and descriptors confirm our theoretical findings. Notation: We use capital letters to denote scalar random variables X and X to denote vector random variables, corresponding small letters x and x to denote the realisations of scalar and vector random variables, respectively. We use X ~pX(x) or simply X ~p(x) to indicate that a random variable X is distributed according to pX(x). N(μ, σ 2 X ) stands for the Gaussian distribution with mean μ and variance σ2 X . B(L, Pb) denotes the binomial distribution with sequence length L and probability of success Pb. ║.║denotes the Euclidean vector norm and Q(.) stands for the Q-function. D(.║.) denotes the divergence and E{.} denotes the expectation.

Secure Representation of Images Using Multi-layer Compression

We analyze the privacy preservation capabilities of a previously introduced multi-stage image representation framework where blocks of images with similar statistics are decomposed into different codebooks (dictionaries). There it was shown that at very low rate regimes, the method is capable of compressing images that come from the same family with results superior to those of the JPEG2000 codec. We consider two different elements to be added to the discussed approach to achieve a joint compression-encryption framework. The first visual scrambling is the random projections were the random matrix is kept secret between the encryption and decryption sides. We show that for the second approach, scrambling in the DCT domain, we can even slightly increase the compression performance of the multi-layer approach while making it safe against de-scrambling attacks. The experiments were carried out on the ExtendedYaleB database of facial images.

Sparse ternary codes for similarity search have higher coding gain than dense binary codes

This paper addresses the problem of Approximate Nearest Neighbor (ANN) search in pattern recognition where feature vectors in a database are encoded as compact codes in order to speed-up the similarity search in large-scale databases. Considering the ANN problem from an information-theoretic perspective, we interpret it as an encoding, which maps the original feature vectors to a less entropic sparse representation while requiring them to be as informative as possible. We then define the coding gain for ANN search using information-theoretic measures. We next show that the classical approach to this problem, which consists of binarization of the projected vectors is sub-optimal. Instead, a properly designed ternary encoding achieves higher coding gains and lower complexity.

Robust human face recognition based on locality preserving sparse over complete block approximation

Compressive Sensing (CS) has become one of the standard methods in face recognition due to the success of the family of Sparse Representation based Classification (SRC) algorithms. However it has been shown that in some cases, the locality of the dictionary codewords is more essential than the sparsity. Also sparse coding does not guarantee to be local which could lead to an unstable solution. We therefore consider the statistically optimal aspects of encoding that guarantee the best approximation of the query image to a dictionary that incorporates varying acquisition conditions. We focus on the investigation, analysis and experimental validation of the best robust classifier/predictor and consider frontal face image variability induced by noise, lighting, expression, pose, etc.. We compare two image representations using a pixel-wise approximation and an overcomplete block-wise approximation with two types of sparsity priors. In the first type we consider all samples from a single subject and in the second type we consider all samples from all subjects. The experiments on a publicly available dataset using low resolution images showed that several per subject sample sparsity prior approximations are as good as the results presented from SCR and that our simple overcomplete block-wise approximation provides superior performance in comparison to the SRC and WSRC algorithm.

Local active content fingerprinting: Optimal solution under linear modulation

This papers presents an analysis on Active Content Fingerprint (aCFP) for local (patch based) image descriptors. A generalization is proposed, the reduction of the aCFP with linear modulation to a constrained projection problem is shown and the optimal solution is given. The constrained projection problem addresses the linear modulation by a constraint on the properties of the resulting local descriptor. A computer simulation using local image patches, extracted from publicly available data sets is provided, demonstrating the advantages under several signal processing distortions.

Supervised Transform Learning for Face Recognition

In this paper we investigate transform learning and apply it to face recognition problem. The focus is to find a transformation matrix that transforms the signal into a robust to noise, discriminative and compact representation. We propose a method that finds an optimal transform under the above constrains. The non-sparse variant of the presented method has a closed form solution whereas the sparse one may be formulated as a solution to a sparsity regularized problem. In addition we give a generalized version of the proposed problem and we propose a prior on the data distribution across the dimensions in the transform domain.

Robust Face Recognition by Group Sparse Representation That Uses Samples from List of Subjects

In this paper we consider group sparsity for robust face recognition. We propose a model for inducing group sparsity with no constraints on the definition of the structure of the group, coupled with locality constrained regularization. We formulate the problem as bounded distance regularized L 2 norm minimization with group sparsity inducing, non-convex constrains. We apply convex relaxation and a branch and bound strategy to find an approximation to the original problem. The empirical results confirm that with this approach of deploying a very simple non-overlapping group structure we outperform several state-of-the-art sparse coding based image classification methods.

Content identification: binary content fingerprinting versus binary content encoding

In this work, we address the problem of content identification. We consider content identification as a special case of multiclass classification. The conventional approach towards identification is based on content fingerprinting where a short binary content description known as a fingerprint is extracted from the content. We propose an alternative solution based on elements of machine learning theory and digital communications. Similar to binary content fingerprinting, binary content representation is generated based on a set of trained binary classifiers. We consider several training/encoding strategies and demonstrate that the proposed system can achieve the upper theoretical performance limits of content identification. The experimental results were carried out both on a synthetic dataset with different parameters and the FAMOS dataset of microstructures from consumer packages.

Joint learning of local fingerprint and content modulation

This paper proposes learning a linear map with local content modulation for robust content fingerprinting. The goal is to estimate a data adapted linear map that provides bounded modulation distortion and features with targeted properties. A novel problem formulation is presented that jointly addresses the fingerprint learning and the content modulation. A solution by iterative alternating algorithm is proposed. The algorithm alternates between liner map update step and linear modulation estimate step. Global optimal solutions for the respective iterative steps are proposed, resulting in convergent algorithm with locally optimal solution. A computer simulation using local image patches, extracted from publicly available data set is provided. The advantages under additive white Gaussian noise (AWGN), lossy JPEG compression and projective geometrical transform distortions are demonstrated.