Visual dictionary

Abstract: We study joint learning of Convolutional Neural Network (CNN) and Transformer for vision-language pre-training (VLPT) which aims to learn cross-modal alignments from millions of image-text pairs. State-of-the-art approaches extract salient image regions and align regions with words step-by-step. As region-based visual features usually represent parts of an image, it is challenging for existing vision-language models to fully understand the semantics from paired natural languages. In this paper, we propose SOHO to "Seeing Out of tHe bOx" that takes a whole image as input, and learns vision-language representation in an end-to-end manner. SOHO does not require bounding box annotations which enables inference 10 times faster than region-based approaches. In particular, SOHO learns to extract comprehensive yet compact image features through a visual dictionary (VD) that facilitates cross-modal understanding. VD is designed to represent consistent visual abstractions of similar semantics. It is updated on-the-fly and utilized in our proposed pre-training task Masked Visual Modeling (MVM). We conduct experiments on four well-established vision-language tasks by following standard VLPT settings. In particular, SOHO achieves absolute gains of 2.0% R@1 score on MSCOCO text retrieval 5k test split, 1.5% accuracy on NLVR2 test-P split, 6.7% accuracy on SNLI-VE test split, respectively.

Abstract: Scene classification has been proved to be an effective method for high spatial resolution (HSR) remote sensing image semantic interpretation. The probabilistic topic model (PTM) has been successfully applied to natural scenes by utilizing a single feature (e.g., the spectral feature); however, it is inadequate for HSR images due to the complex structure of the land-cover classes. Although several studies have investigated techniques that combine multiple features, the different features are usually quantized after simple concatenation (CAT-PTM). Unfortunately, due to the inadequate fusion capacity of $\boldsymbol{k}$ -means clustering, the words of the visual dictionary obtained by CAT-PTM are highly correlated. In this paper, a semantic allocation level (SAL) multifeature fusion strategy based on PTM, namely, SAL-PTM (SAL-pLSA and SAL-LDA) for HSR imagery is proposed. In SAL-PTM: 1) the complementary spectral, texture, and scale-invariant-feature-transform features are effectively combined; 2) the three features are extracted and quantized separately by $\boldsymbol{k}$ -means clustering, which can provide appropriate low-level feature descriptions for the semantic representations; and 3)the latent semantic allocations of the three features are captured separately by PTM, which follows the core idea of PTM-based scene classification. The probabilistic latent semantic analysis (pLSA) and latent Dirichlet allocation (LDA) models were compared to test the effect of different PTMs for HSR imagery. A U.S. Geological Survey data set and the UC Merced data set were utilized to evaluate SAL-PTM in comparison with the conventional methods. The experimental results confirmed that SAL-PTM is superior to the single-feature methods and CAT-PTM in the scene classification of HSR imagery.

Abstract: Computer vision-based food recognition could be used to estimate a meal's carbohydrate content for diabetic patients. This study proposes a methodology for automatic food recognition, based on the bag-of-features (BoF) model. An extensive technical investigation was conducted for the identification and optimization of the best performing components involved in the BoF architecture, as well as the estimation of the corresponding parameters. For the design and evaluation of the prototype system, a visual dataset with nearly 5000 food images was created and organized into 11 classes. The optimized system computes dense local features, using the scale-invariant feature transform on the HSV color space, builds a visual dictionary of 10000 visual words by using the hierarchical k-means clustering and finally classifies the food images with a linear support vector machine classifier. The system achieved classification accuracy of the order of 78%, thus proving the feasibility of the proposed approach in a very challenging image dataset.