Deep convolutional networks can solve various complex tasks in the field of image processing. However, adversarial attacks have been shown to have the ability of fooling deep learning models. Adversarial training is one commonly used strategy to improve the robustness of deep learning models against adversarial examples, which is performed by incorporating adversarial examples into the training process. Traditionally, during this process, cross-entropy loss is used as the loss function. In order to improve the robustness of deep learning models against adversarial examples, we propose in this paper two new methods of adversarial training by applying the principle of Maximal Coding Rate Reduction (MCR2). We evaluate the performance of different adversarial training methods by comparing the clean accuracy and adversarial accuracy. It is shown that adversarial training with the MCR2 loss function yields a more robust network than the traditional adversarial training method. In our experiments, adversarial accuracies are improved by up to 10%. The two loss functions are discussed by using a model.

Infrared object detection is one of the critical technologies for maritime search and rescue. However, it is still challenging due to the strong background clutter interference and the lack of small object information. We proposed a temporal-spatial pyramid network for infrared maritime object detection. We proposed a nested temporal pyramid to represent the temporal features through motion differences maps and energy accumulation maps to distinguish the wave clutter and objects. We proposed a dense spatial pyramid to learn the spatial features and the differences between temporal maps and then to clarify and locate objects. For training, we designed a scale-related composite loss function with correlated location description and weighted confidence loss. Finally, based on the ablation and comparison experiments, the proposed method performs better on maritime infrared sequences.

We present a machine learning approach that uses a custom Convolutional Neural Network (CNN) for estimating the depth of water pools from multispectral drone imagery. Using drones to obtain this information offers a cheaper, timely, and more accurate solution compared to alternative methods, such as manual inspection. This information, in turn, represents an asset to identify potential breeding sites of mosquito larvae, which grow only in shallow water pools. As a significant part of the world's population is affected by mosquito-borne viral infections, including Dengue and Zika, identifying mosquito breeding sites is key to control their spread. Experiments with 5-band drone imagery show that our CNN-based approach is able to measure shallow water depths accurately up to a root mean square error of less than 0.5 cm, outperforming state-of-the-art Random Forest methods and empirical approaches.

Mosquitoes spread disases such as Dengue and Zika that affect a significant portion of the world population. One approach to hamper the spread of the disases is to identify the mosquitoes’ breeding places. Recent studies use drones to detect breeding sites, due to their low cost and flexibility. In this paper, we investigate the applicability of drone-based multi-spectral imagery and mmWave radios to discover breeding habitats. Our approach is based on the detection of water bodies. We introduce our Faster R-CNN-MSWD, an extended version of the Faster R-CNN object detection network, which can be used to identify water retention areas in both urban and rural settings using multi-spectral images. We also show promising results for estimating extreme shallow water depth using drone-based multi-spectral images. Further, we present an approach to detect water with mmWave radios from drones. Finally, we emphasize the importance of fusing the data of the two sensors and outline future research directions.

We present a representation learning framework for financial time series forecasting. One challenge of using deep learning models for finance forecasting is the shortage of available training data when using small datasets. Direct trend classification using deep neural networks trained on small datasets is susceptible to the overfitting problem. In this paper, we propose to first learn compact representations from time series data, then use the learned representations to train a simpler model for predicting time series movements. We consider a class-conditioned latent variable model. We train an encoder network to maximize the mutual information between the latent variables and the trend information conditioned on the encoded observed variables. We show that conditional mutual information maximization can be approximated by a contrastive loss. Then, the problem is transformed into a classification task of determining whether two encoded representations are sampled from the same class or not. This is equivalent to performing pairwise comparisons of the training datapoints, and thus, improves the generalization ability of the encoder network. We use deep autoregressive models as our encoder to capture long-term dependencies of the sequence data. Empirical experiments indicate that our proposed method has the potential to advance state-of-the-art performance.

Autoencoders and their variations provide unsupervised models for learning low-dimensional representations for downstream tasks. Without proper regularization, autoencoder models are susceptible to the overfitting problem and the so-called posterior collapse phenomenon. In this paper, we introduce a quantization-based regularizer in the bottleneck stage of autoencoder models to learn meaningful latent representations. We combine both perspectives of Vector Quantized-Variational AutoEncoders (VQ-VAE) and classical denoising regularization methods of neural networks. We interpret quantizers as regularizers that constrain latent representations while fostering a similarity-preserving mapping at the encoder. Before quantization, we impose noise on the latent codes and use a Bayesian estimator to optimize the quantizer-based representation. The introduced bottleneck Bayesian estimator outputs the posterior mean of the centroids to the decoder, and thus, is performing soft quantization of the noisy latent codes. We show that our proposed regularization method results in improved latent representations for both supervised learning and clustering downstream tasks when compared to autoencoders using other bottleneck structures.

Fractional-pel accurate motion is widely used in video coding. For subband coding, fractional-pel accuracy is challenging since it is difficult to handle the complex motion field with temporal transforms. In our previous work, we designed integer accurate motion-adaptive transforms (MAT) which can transform integer accurate motion-connected coefficients. In this paper, we extend the integer MAT to fractional-pel accuracy. The integer MAT allows only one reference coefficient to be the lowhand coefficient. In this paper, we design the transform such that it permits multiple references and generates multiple low-band coefficients. In addition, our fractional-pel MAT can incorporate a general interpolation filter into the basis vector, such that the highband coefficient produced by the transform is the same as the prediction error from the interpolation filter. The fractional-pel MAT is always orthonormal. Thus, the energy is preserved by the transform. We compare the proposed fractional-pel MAT, the integer MAT, and the half-pel motion-compensated orthogonal transform (MCOT), while HEVC intra coding is used to encode the temporal subbands. The experimental results show that the proposed fractional-pel MAT outperforms the integer MAT and the half-pel MCOT. The gain achieved by the proposed MAT over the integer MAT can reach up to 1 dB in PSNR.

Vector-Quantized Variational Autoencoders (VQ-VAE)[1] provide an unsupervised model for learning discrete representations by combining vector quantization and autoencoders. In this paper, we study the use of VQ-VAE for representation learning of downstream tasks, such as image retrieval. First, we describe the VQ-VAE in the context of an information-theoretic framework. Then, we show that the regularization effect on the learned representation is determined by the size of the embedded codebook before the training. As a result, we introduce a hyperparameter to balance the strength of the vector quantizer and the reconstruction error. By tuning the hyperparameter, the embedded bottleneck quantizer is used as a regularizer that forces the output of the encoder to share a constrained coding space. With that, the learned latent features better preserve the similarity relations of the data space. Finally, we incorporate the product quantizer into the bottleneck stage of VQ-VAE and use it as an end-to-end unsupervised learning model for image retrieval tasks. The product quantizer has the advantage of generating large and structured codebooks. Fast retrieval can be achieved by using lookup tables that store the distance between any pair of sub-codewords. State-of-the-art retrieval results are achieved by the proposed codebooks. 

This paper considers the problem of compression for similarity identification. Unlike classic compression problems, the focus is not on reconstructing the original data. Instead, compression is determined by the reliability of answering given queries. The problem is characterized by the identification rate of a source which is the minimum compression rate which allows reliable answers for a given similarity threshold. In this work, we investigate the component-based quadratic similarity identification for multivariate Gaussian sources. The decorrelated original data is processed by a distinct D- A dmissible system for each component. For a special case, we characterize the component-based identification rate for a correlated Gaussian source. Furthermore, we derived the optimal bit allocation for a given total rate constraint.