Об одном подходе к последовательному иерархическому распознаванию изображений
In this paper we explore an application of the pyramid HOG (Histograms of Oriented Gradients) features in image recognition problem with small samples. A sequential analysis is used to improve the performance of hierarchical methods. We propose to process the next, more detailed level of pyramid only if the decision at the current level is unreliable. The Chow’s reject option of comparison of the posterior probability with a fixed threshold is used to verify recognition reliability. The posterior probability is estimated for the homogeneity-testing probabilistic neural network classifier on the basis of its relation with the Bayesian decision. Experimental results in face recognition are presented. It is shown that the proposed approach allows to increase the recognition performance in 2–4 times in comparison with conventional classification of pyramid HOGs.
The article is devoted to pattern recognition task with the database containing small number of samples per class. By mapping of local continuous feature vectors to a discrete range, this problem is reduced to statistical classification of a set of discrete finite patterns. It is demonstrated that Bayesian decision under the assumption that probability distributions can be estimated using the Parzen kernel and the Gaussian window with a fixed variance for all the classes, implemented in the PNN, is not optimal in the classification of a set of patterns. We presented here the novel modification of the PNN with homogeneity testing which gives an optimal solution of the latter task under the same assumption about probability densities. By exploiting the discrete nature of patterns our modification prevents the well-known drawbacks of the memory-based approach implemented in both the PNN and the PNN with homogeneity testing, namely, low classification speed and high requirements to the memory usage. Our modification only requires the storage and processing of the histograms of input and training samples. We present the results of an experimental study in two practically important tasks: 1) the problem of Russian text authorship attribution with character n-grams features; and 2) face recognition with well-known datasets (AT&T, FERET and JAFFE) and comparison of color- and gradient-orientation histograms. Our results support the statement that the proposed network provides better accuracy (1-7%) and is much more resistant to change of the smoothing parameter of Gaussian kernel function in comparison with the original PNN.
Probabilistic neural network (PNN) is the well-known instance-based learning algorithm, which is widely used in various pattern classification and regression tasks, if rather small number of instances for each class is available. The known disadvantage of this network is its insufficient classification computational complexity. The common way to overcome this drawback is the reduction techniques with selection of the most typical instances. Such approach causes the shifting of the estimates of the class probability distribution, and, in turn, the decrease of the classification accuracy. In this paper we examine another possible solution by replacing the Gaussian window and the Parzen kernel to the orthogonal series Fejér kernel and using the naïve assumption about independence of features. It is shown, that our approach makes it possible to achieve much better runtime complexity in comparison with either original PNN or its modification with the preliminary clustering of the training set.
This paper addresses the problem of insufficient performance of statistical classification with the medium-sized database (thousands of classes). Each object is represented as a sequence of independent segments. Each segment is defined as a random sample of independent features with the distribution of multivariate exponential type. To increase the speed of the optimal Kullback-Leibler minimum information discrimination principle, we apply the clustering of the training set and an approximate nearest neighbor search of the input object in a set of cluster medoids. By using the asymptotic properties of the Kullback-Leibler divergence, we propose the maximal likelihood search procedure. In this method the medoid to check is selected from the cluster with the maximal joint density (likelihood) of the distances to the previously checked medoids. Experimental results in image recognition with artificially generated dataset and Essex facial database prove that the proposed approach is much more effective, than an exhaustive search and the known approximate nearest neighbor methods from FLANN and NonMetricSpace libraries.
The usage of the probabilistic neural network with homogeneity testing is proposed in image recognition problem. This decision is shown to be optimal in Bayesian terms if the task is formulated as a statistical testing for homogeneity of query and model images' feature sets. The problem of the lack of computing efficiency with many classes and large dimensions of feature set is discovered. The possibility of its overcoming in the case of discrete features is explored by synthesizing the novel recognition criterion with the comparison of the histograms of query and model images. It is shown that a particular case of this criterion is the nearest neighbor rule with popular measures of similarity, namely, chi-square distance and Jensen-Shannon divergence. The results of experimental research in a problem of face recognition with widely used databases (AT&T, JAFFE) are presented. The proposed approach is demonstrated to achieve better recognition accuracy in comparison with conventional solution with reduction the recognition task to the statistical classification.