Le Lv

Deep Reinforcement Learning With Visual Attention for Vehicle Classification

Automatic vehicle classification is crucial to intelligent transportation system, especially for vehicle-tracking by police. Due to the complex lighting and image capture conditions, image-based vehicle classification in real-world environments is still a challenging task and the performance is far from being satisfactory. However, owing to the mechanism of visual attention, the human vision system shows remarkable capability compared with the computer vision system, especially in distinguishing nuances processing. Inspired by this mechanism, we propose a convolutional neural network (CNN) model of visual attention for image classification. A visual attention-based image processing module is used to highlight one part of an image and weaken the others, generating a focused image. Then the focused image is input into the CNN to be classified. According to the classification probability distribution, we compute the information entropy to guide a reinforcement learning agent to achieve a better policy for image classification to select the key parts of an image. Systematic experiments on a surveillance-nature dataset which contains images captured by surveillance cameras in the front view, demonstrate that the proposed model is more competitive than the large-scale CNN in vehicle classification tasks.

A Semi-Supervised Predictive Sparse Decomposition Based on Task-Driven Dictionary Learning

In feature learning field, many methods are inspired by advances in neuroscience. Among them, neural network and sparse coding have been broadly studied. Predictive sparse decomposition (PSD) is a practical variant of these two methods. It trains a neural network to estimate the sparse codes. After training, the neural network is fine-tuned to achieve higher performance on object recognition tasks. It is widely believed that introducing discriminative information can make the features more useful for classification task. Hence, in this work, we propose applying the task-driven dictionary learning framework to the PSD and demonstrate that this new model can be optimized by the stochastic gradient descent (SGD) algorithm. Before our work, the semi-supervised auto-encoder framework has already been proposed to guide neural network to extract discriminative representations. But it does not improve the classification performance of neural network. In the experiments, we compare the proposed method with the semi-supervised auto-encoder method. The performance of PSD is used as the baseline for these two methods. On the MNIST and USPS datasets, our method can generate more discriminative and predictable sparse codes than other methods. Furthermore, the recognition accuracy of neural network can be improved.

Multi-task learning for dangerous object detection in autonomous driving☆

Abstract Recently, autonomous driving has been extensively studied and has shown considerable promise. Vision-based dangerous object detection is a crucial technology of autonomous driving. In previous work, dangerous object detection is generally formulated as a typical object detection problem and a distance-based danger assessment problem, separately. These two problems are usually dealt with using two independent models. In fact, vision-based object detection and distance prediction present prominent visual relationship. The objects with different distance to the camera have different attributes (pose, size and definition), which are very worthy to be exploited for dangerous object detection. However, these characteristics are usually ignored in previous work. In this paper, we propose a novel multi-task learning (MTL) method to jointly model object detection and distance prediction with a Cartesian product-based multi-task combination strategy. Furthermore, we mathematically prove that the proposed Cartesian product-based combination strategy is more optimal than the linear multi-task combination strategy that is usually used in MTL models, when the multi-task itself is not independent. Systematic experiments show that the proposed approach consistently achieves better object detection and distance prediction performances compared to both the single-task and multi-task dangerous object detection methods.

A visual attention based convolutional neural network for image classification

This paper presents a visual attention based convolutional neural network (CNN) to solve the image classification problem in the real complex world scene. The presented method can simulate the process of recognizing objects and find the area of interest which is related with the task. Compared with the CNN method in image classification, the model is proficient in fine-grained classification problem and has a better robustness due to its mechanism of multi-glance and visual attention. We evaluate the model on vehicle dataset, where its performance exceeds CNN baseline on image classification.

Graphene size-dependent modulation of graphene frameworks contributing to the superior thermal conductivity of epoxy composites

Vacuum filtration is a highly effective and easy scale-up approach and has been widely used to fabricate graphene monoliths, such as graphene paper and graphene frameworks for various applications. In general, the microstructure of filtrated monoliths exhibits layer-by-layer stacking of graphene sheets due to the directional flow-induced assembly process. In this work, we found that the horizontally oriented structure of filtrated graphene frameworks can be modulated to an approximately isotropic arrangement by lowering the lateral size of graphene sheets. This size-dependent microstructure transition from anisotropic to isotropic was further confirmed by measuring the in- and through-plane thermal conductivity of the graphene/epoxy composites with different arrangements of graphene frameworks as a filler. Optimally, we obtained an epoxy composite embedded with a quasi-isotropic graphene framework (QIGF) by a simple two-step process: vacuum filtration of small graphene sheets to obtain the framework followed by the infiltration of epoxy resin. Based on the interconnected graphene sheets with an approximately isotropic arrangement, QIGF provides heat channels of graphene–graphene along both the in- and through-plane directions within epoxy. With a low graphene loading of 5.5 wt%, QIGF/epoxy (QIGF/EP) presents in- and through-plane thermal conductivities of 10.0 and 5.4 W mK−1, respectively, which are equivalent to ∼55 and 29 times higher than those of neat epoxy. As compared to the current graphene/epoxy composites prepared by various methods, our QIGF/EP has the highest thermal conductivity value with this level of filler loading. Our findings provide an insight for the development of polymer composites for thermal management applications in industry.

Convolutional fitted Q iteration for vision-based control problems

In this paper a deep reinforcement learning (DRL) method is proposed to solve the control problem which takes raw image pixels as input states. A convolutional neural network (CNN) is used to approximate Q functions, termed as Q-CNN. A pretrained network, which is the result of a classification challenge on a vast set of natural images, initializes the parameters of Q-CNN. Such initialization assigns Q-CNN with the features of image representation, so it is more concentrated on the control tasks. The weights are tuned under the scheme of fitted Q iteration (FQI), which is an offline reinforcement learning method with the stable convergence property. To demonstrate the performance, a modified Food-Poison problem is simulated. The agent determines its movements based on its forward view. In the end the algorithm successfully learns a satisfied policy which has better performance than the results of previous researches.

ADP with MCTS algorithm for Gomoku

Inspired by the core idea of AlphaGo, we combine a neural network, which is trained by Adaptive Dynamic Programming (ADP), with Monte Carlo Tree Search (MCTS) algorithm for Gomoku. MCTS algorithm is based on Monte Carlo simulation method, which goes through lots of simulations and generates a game search tree. We rollout it and search the outcomes of the leaf nodes in the tree. As a result, we obtain the MCTS winning rate. The ADP and MCTS methods are used to estimate the winning rates respectively. We weight the two winning rates to select the action position with the maximum one. Experiment result shows that this method can effectively eliminate the neural network evaluation functions “short-sighted” defect. With our proposed method, the games final prediction result is more accurate, and it outperforms the Gomoku with ADP algorithm.

Image clustering based on deep sparse representations

Currently, the supervised trained deep neural networks (DNNs) have been successfully applied in several image classification tasks. However, how to extract powerful data representations and discover semantic concepts from unlabeled data is a more practical issue. Unsupervised feature learning methods aim at extracting abstract representations from unlabeled data. Large amount of research works illustrate that these representations can be directly used in the supervised tasks. However, due to the high dimensionality of these representations, it is difficult to discover the categorical concepts among them in an unsupervised way. In this paper, we propose combining the winner-take-all autoencoder with the bipartite graph partitioning algorithm to cluster unlabeled image data. The winner-take-all autoencoder can learn the additive sparse representations. By the experiments, we present the properties of the sparse representations. The bipartite graph partitioning can take full advantage of them and generate semantic clusters. We discover that the confident instances in each cluster are well discriminated. Based on the initial clustering result, we further train a support vector machine (SVM) to refine the clusters. Our method can discover the categorical concepts rapidly and the experiment shows that the clustering performance of our method is good.