Compressing deep convolutional neural networks in visual emotion recognition
It has been shown that the activations invoked by an image within the top layers of a large convolutional neural network provide a high-level descriptor of the visual content of the image. In this paper, we investigate the use of such descriptors (neural codes) within the image retrieval application. In the experiments with several standard retrieval benchmarks, we establish that neural codes perform competitively even when the convolutional neural network has been trained for an unrelated classification task (e.g. Image-Net). We also evaluate the improvement in the retrieval performance of neural codes, when the network is retrained on a dataset of images that are similar to images encountered at test time. We further evaluate the performance of the compressed neural codes and show that a simple PCA compression provides very good short codes that give state-of-the-art accuracy on a number of datasets. In general, neural codes turn out to be much more resilient to such compression in comparison other state-of-the-art descriptors. Finally, we show that discriminative dimensionality reduction trained on a dataset of pairs of matched photographs improves the performance of PCA-compressed neural codes even further. Overall, our quantitative experiments demonstrate the promise of neural codes as visual descriptors for image retrieval.
A new public dataset of traffic sign images is presented. The dataset is intended for training and testing the algorithms of traffic sign recognition. We describe the dataset structure and guidelines for working with the dataset, comparing it with the previously published traffic sign datasets. The evaluation of modern detection and classification algorithms conducted using the proposed dataset has shown that existing methods of recognition of a wide class of traffic signs do not achieve the accuracy and completeness required for a number of applications.
In this paper we focus on the image recognition problem in the case of small sample size based on the nearest neighbor rule and matching of high-dimensional feature vectors extracted with the deep convolutional neural network. We propose the novel recognition algorithm based on the maximum likelihood method for the joint density of dissimilarities between an observed image and available instances in the training set. This likelihood is estimated using the known asymptotically normally distribution of the Jensen-Shannon divergence between image features, if the latter can be treated as the probability density estimates. This asymptotic behavior is in agreement with the well-known experimental estimates of distributions of dissimilarity distances between high-dimensional vectors. The experimental study in unconstrained face recognition for the LFW (Labeled Faces in the Wild) and YTF (YouTube Faces) datasets demonstrated, that the proposed approach makes it possible to increase the recognition accuracy at 1-5% when compared with conventional classifiers.
In this paper we compare the Russian National Corpus to a larger Russian web corpus composed in 2014; the assumption behind our work is that the National corpus, being limited by the texts it contains and their proportions, presents lexical contexts (and thus meanings) which are different from those found ‘in the wild’ or in a language in use.
To do such a comparison, we used both corpora as training sets to learn vector word representations and found the nearest neighbors or associates for all top-frequency nominal lexical units. Then the difference between these two neighbor sets for each word was calculated using the Jaccard similarity coefficient. The resulting value is the measure of how much the meaning of a given word is different in the language of web pages from the Russian language in the National corpus. About 15% of words were found to acquire completely new neighbors in the web corpus.
In this paper, the methodology of research is described and implications for Russian National Corpus are proposed. All experimental data are available online.
An exhaustive search of all classes in pattern recognition methods cannot be implemented in real-time, if the database contains a large number of classes. In this paper we introduce a novel probabilistic approximate nearest-neighbor (NN) method. Despite the most of known fast approximate NN algorithms, our method is not heuristic. The joint probabilistic densities (likelihoods) of the distances to previously checked reference objects are estimated for each class. The next reference instance is selected from the class with the maximal likelihood. To deal with the quadratic memory requirement of this approach, we propose its modification, which processes the distances from all instances to a small set of pivots chosen with the farthest-first traversal. Experimental study in face recognition with the histograms of oriented gradients and the deep neural network-based image features shows that the proposed method is much faster than the known approximate NN algorithms for medium databases.