The polarization-based attitude and heading reference system (PAHRS) consisting of inertial navigation system (INS) and polarization sensor (PS) offers an effective solution for attitude and heading determination in the case of global navigation satellite system (GNSS) signal degradation. Its performance depends largely on the state estimation accuracy. In the existing work, the Kalman filtering (KF), as a low complexity scheme, is employed to achieve the state estimation of PAHRS. However, in practice, the performance of PAHRS could be affected by weather conditions and the maneuvering state of the vehicle. The accurate noise statistics of INS and PS is often encountered, which leads to a degradation of PAHRS. To improve the adaptability and accuracy of the system, in this article, we conduct a KF flow-based deep neural networks (KFDNNs), a real-time state estimator that learns from PS and INS data to carry out Kalman filter. In the constructed KFDNNs, deep neural networks (DNNs) are inserted into the flow of the KF to learn the optimal Kalman gain from PAHRS data, which we can retain data efficiency and interpretability of the classic algorithm while circumvents the dependency of the KF on knowledge of the noise statistics. Moreover, a two-stage training strategy consisting of warm-up training stage and task-oriented training stage is presented for the KFDNNs, which mitigate the gradient explosion caused by the unstable random initialization of KFDNNs while improve the flexibility of sequence length selection. Finally, the simulation and vehicle test are carried to verify the performance of PAHRS. The experimental results confirm the KFDNNs outperforms KF-based INS/PS method especially in complex weather and maneuvering scenarios.

Underwater images suffer from blur, low contrast, color distortion and loss of details. Recently, since polarized images can provide more information, many underwater image enhancement(UIE) methods based on polarized images are proposed. Most existing polarized methods usually enhance images by physical techniques based on some priors or assumptions. However, the priors or assumptions are not always reliable in the real scene, which limits their performance. Besides, existing deep learning methods fuse four polarized images at different angles (I0–3) into a clear visible light image, which only use simple fusion methods and do not preserve image information well. In order to avoid the limitation of priors and obtain the rich information contained in polarized images more reasonable, we propose a novel framework that using the quality maps to guide the fusion of the four images to enhance images, called Quality-Guided Enhancement Network(QGE-Net). Specifically, QGE-Net can be divided into Quality Evaluator and Enhancement Module. A loss called SROCC Rank Loss is proposed to train Quality Evaluator, which can improve the ability of Quality Evaluator to obtain differential information from different images in the same scene. In Enhancement Module, a module called Multi-Scale Quality Guidance Fusion(MQGF) is proposed to fuse the features of polarized images under the guidance of the quality map to make full use of the information of the input image. Extensive experimental results show that the proposed method achieves superior performance against state-of-the-art methods on both synthetic and real data.

Underwater images suffer from blur, low contrast, color distortion and loss of details. Recently, since polarized images can provide more information, many underwater image enhancement(UIE) methods based on polarized images are proposed. Most existing polarized methods usually enhance images by physical techniques based on some priors or assumptions. However, the priors or assumptions are not always reliable in the real scene, which limits their performance. Besides, existing deep learning methods fuse four polarized images at different angles ($\mathbf{I}_{0-3}$) into a clear visible light image, which only use simple fusion methods and do not preserve image information well. In order to avoid the limitation of priors and obtain the rich information contained in polarized images more reasonable, we propose a novel framework that using the quality maps to guide the fusion of the four images to enhance images, called QualityGuided Enhancement Network(QGE-Net). Specifically, QGE-Net can be divided into Quality Evaluator and Enhancement Module. A loss called SROCC Rank Loss is proposed to train Quality Evaluator, which can improve the ability of Quality Evaluator to obtain differential information from different images in the same scene. In Enhancement Module, a module called MultiScale Quality Guidance Fusion(MQGF) is proposed to fuse the features of polarized images under the guidance of the quality map to make full use of the information of the input image. Extensive experimental results show that the proposed method achieves superior performance against state-of-the-art methods on both synthetic and real data.

Images captured in underwater environments often suffer from color distortion, low contrast, and reduced visual quality. Most existing methods solve underwater image enhancement (UIE) by applying supervised training on synthetic images or pseudo references. However, the synthetic paired data fail to accurately replicate real-world data due to the inherent differences, and the quantity and quality of pseudo references are limited, which seriously reduces the generalization ability and performance of the model when testing on real underwater images. In contrast, unsupervised-based method is not constrained by paired data, which is more robust and potentially more promising for practical applications. Nevertheless, existing unsupervised-based methods cannot effectively constrain the network to train a model that can adapt to various degradation. Inspired by the fact that people often resolve problems from opposing but complementary perspectives, we maintain that there is implicit cooperation between the removal and generation of water layers, as they can constrain and promote each other at the same time. Based on the above analysis, a new unsupervised-based UIE framework that jointly learns water layer generation and removal based on disentangled representations is proposed. Specifically, we propose a bidirectional disentangling network in which each unidirectional network contains a loop consisting of water layer removal and generation, and restricts the image to remain consistent after a loop. Meanwhile, a novel double-order contrastive loss is proposed to improve the ability of disentanglement by utilizing the joint implicit constraint of first-order features and second-order features. Extensive experimental results demonstrate that the model outperforms the state-of-the-art methods in both qualitative and quantitative evaluation with a relatively high processing speed. The experimental results of the ablation study demonstrate the usefulness of the various components.

Different from single image dehazing, polarized images can record richer scene information, and thus polarized images-based dehazing has attracted tremendous attention recently. Existing polarized images-based dehazing methods usually propose some priors or assumptions to calculate the degree of polarization or angle of polarization for dehazing. Although these methods have made significant progress, the priors or assumptions are not always reliable in the real scene, which limits their performance. In addition, most existing methods only consider pixel-level difference information and ignore extracting more effective information from themselves. Based on the above analysis, we propose a novel framework that transforms the polarized images-based dehazing problem into a multi-image fusion problem without any assumption based on the fact that the source of the polarized image is from the same scene but contains different scene information. Specifically, we first pre-train a generic dehazing physical model to obtain an intermediate result that serves as a reference image due to a clearer structure. Then the reference image is used to guide the features fusion to extract more effective features from the input polarized images themselves. Extensive experimental results show that the proposed method achieves superior performance against state-of-the-art methods on both synthetic and real data.

In this article, a novel control algorithm is developed for a class of nonlinear stochastic systems subject to multiple disturbances, including exogenous dynamic disturbance and general non-Gaussian noise. An observer is designed to estimate the exogenous disturbance, and then the disturbance compensation is incorporated into a feedback control strategy for the non-Gaussian system. Considering the ability of entropy in randomness quantification, a performance index is established based on the generalized entropy optimization principle. Furthermore, it is adjusted to be available for the controller solution, which also solves the coupling between two kinds of disturbances. On this basis, the optimal controller is provided in a recursive way, with which the closed-loop stability and good antidisturbance ability can be guaranteed simultaneously. Compared with the existing studies on the non-Gaussian stochastic systems, the proposed control algorithm has merits in multiple disturbances decoupling and enhanced antidisturbance performance. Finally, a simulation example is given to demonstrate the effectiveness of theoretical results.

Fatigue driving is one of the main causes of traffic accidents. For real-world driver fatigue detection, the large pose deformations exhibited by the captured global face significantly increase the difficulty of extracting effective features. Furthermore, previous fatigue detection methods have not achieved desired results in distinguishing actions with similar appearance, such as yawning and speaking. In this article, we propose a multi-granularity Deep Convolutional Model based on feature Recalibration and Fusion for driver fatigue detection (RF-DCM). Our deep model leverages cues from partial faces to alleviate the pose variations and obtains robust feature representations from both the global face and different local parts. The core innovative techniques are as follows: A multi-granularity extraction sub-network extracts more efficient multi-granularity features while compressing the parameters of the network. In order to match multi-granularity features, a feature rectification sub-network and a feature fusion sub-network are designed to adaptively recalibrate and fuse the multi-granularity features. A long short term memory network is used to explore the relationship among sequence frames to distinguish actions with similar appearances. Extensive experimental results on the public drowsy driver dataset from NTHU Driver Drowsy competition demonstrate significant performance improvements of our model over all published state-of-the-art methods.

Unsupervised learning methods have achieved remarkable performance in monocular depth estimation and camera pose, which mostly solve the multi-task learning problem by using their inner geometry consistency as the self-supervision signal. While most existing approaches mostly adopt the generative model to obtain the depth map prediction, so in the resolution of depth map there is room for improvement. To this end, we present our unsupervised learning architecture based on adversarial learning model, which is used for unsupervised learning of high-resolution single view depth and camera pose. Specifically, we present a multi-scale deep convolutional Generative Adversarial Network (GAN) based learning system, which consists of three networks (pose estimation network PCNN, Generator-D and Discriminator-D for depth map prediction). Furthermore, in order to generate high-resolution depth map, we propose a multi-scale GAN model (MSGAN) to decompose the hard high-quality image generation problem into more manageable sub-problems through a coarse-to-fine process. Then, we modify the overall generation architecture of GAN model by changing the down-sampling and up-sampling components to improve the quality and accuracy of the depth map prediction. Finally, in order to improve the rate of convergence, we use the Least Square Error to increase the penalty for outliers. Detailed quantitative and qualitative evaluations of the proposed framework on the KITTI dataset show that the proposed method provides better results for both pose estimation and depth recovery.