Search

In the production process of laser welding products, visual inspection is usually employed to recognize welding spot locations and diagnose their quality faults. However, commonly used algorithms fail to succeed in both reliability and computational efficiency, especially when applied to assembly line. In this article, a method based on deep learning algorithm and traditional computer vision (TCV) algorithm is proposed, which achieves quality inspection of laser welding spots in the process of battery production. First, compressed U-shape network (CU-net) is proposed to extract welding pads and welding spots. Then, a template-based method is proposed to confirm the validity of each welding spot. Finally, TCV heuristic algorithms are proposed to achieve three error detections, i.e., welding pad placed obliquely, electrode tab placed over highly, and welding spot welded through. Moreover, we build a Welding Spot Quality Inspection Dataset taken from real assembly line. Compared with other pipelines, including U-net, MaskRCNN, and PSPNet, CU-net shows a significant superiority in both processing speed and detection accuracy. The results of template-based method and TCV heuristic algorithms have shown high computational efficiency and ensured inspection accuracy. The inference time of the whole method is less than 100 ms with the implementation on NVIDIA 1060 and Intel i7-6700.

Shot boundary detection (SBD) plays an important role in video understanding, since most recent works take the shot as minimal granularity instead of frames for upstream tasks. However, the large variations of hard-cut and gradual-change transitions within shots significantly limit the performance of SBD. To deal with the variations, we propose a multi-task architecture called Transnet++. Transnet++ disentangles the two types of transition and adopts two separate branches to predict them respectively. Two branches share the same video knowledge space and their results are fused for final prediction. Moreover, we propose a spatial attention module (SAM) to enhance the feature representations which suffers from redundant padding region. Meanwhile, a temporal attention module (TAM) is applied to capture the long-term information of the video for alleviating the over-segmentation problem. Experimental results (91.16% f1-score) on Tencent AVS Dataset demonstrate the effectiveness and superiority of Transnet++ for SBD.

Swarm intelligence algorithms have been widely adopted in solving many highly nonlinear, multimodal problems and have achieved tremendous successes. However, their application on deep neural networks is largely unexplored. On the other hand, deep neural networks, especially convolutional neural network (CNN), have recently achieved breakthroughs in tackling many intractable problems; nevertheless their performance depends heavily on the chosen values of their hyper-parameters, whose fine-tuning is both labor-intensive and time-consuming. In this paper, we propose a novel particle swarm optimization (PSO) variant cPSO-CNN for optimizing the hyper-parameter configuration of architecture-determined CNNs. cPSO-CNN utilizes a confidence function defined by a compound normal distribution to model experts' knowledge on CNN hyper-parameter fine-tunings so as to enhance the canonical PSO's exploration capability. cPSO-CNN also redefines the scalar acceleration coefficients of PSO as vectors to better adapt for the variant ranges of CNN hyper-parameters. Besides, a linear prediction model is adopted for fast ranking the PSO particles to reduce the cost of fitness function evaluation. The experimental results demonstrate that cPSO-CNN performs competitively when compared with several reported algorithms in terms of both CNN hyper-parameter superiority and overall computation cost.

Detecting double JPEG compression is important to forensic experts in identifying the originality and authenticity of images. However, there are some anti-forensic techniques which can evade existing double compression detectors. It is desirable to design a unified approach to address the issues of JPEG forensics and counter-anti-forensics simultaneously, but existing hand-crafted feature based methods and deep learning based methods may fail to satisfy the requirement. In this paper, we present a data-driven approach by using a convolutional neural network (CNN) which takes input from both raw JPEG DCT coefficients and decompressed image pixels. Expert knowledge about JPEG characteristics is incorporated in the CNN design by exploring the intricate relations both within and among DCT subbands and by looking for spatial artifacts both within and among JPEG grids. The CNN is capable of learning deep representations from training data and thus can effectively detect double JPEG compression and its related anti-forensic operations together. The end-to-end CNN that takes into account the information from both DCT domain and spatial domain, shows outstanding performance when compared to prior arts in the experiments. It shows a promising way to address counter-anti-forensic issues without designing specific features for each anti-forensic operation.

In order to improve the ability of a power system with doubly-fed induction generator (DFIG) to suppress low-frequency oscillations, this paper put forward a damping control strategy based on model predictive control (MPC). It is raised that stochastic subspace identification (SSI) is used to identity a power system with DFIG. Compared with other control strategies, configuring MPC can achieve the purpose of system optimization, which is to enhance the stability of a power system with DFIG. This paper uses the 4-machine 2-area system to compare results of MPC and traditional power system stabilizer (PSS) control in terms of damping characteristics, stability and applicability.

With the development of intelligent mobile robots, SLAM, especially V-SLAM, as the basic technology of robot localization and navigation, has the advantages of strong adaptability, high precision and strong intelligence compared with the traditional localization technology. It is widely used in smart devices such as unmanned aerial vehicle, automatic driving and sweeping robots. According to different implementation methods, the visual SLAM is divided into: filter V-SLAM based on probability model, key frame BA-based V-SLAM using nonlinear optimization theory, direct tracking of V-SLAM under the assumption of luminosity invariance, space occupying V-SLAM that focuses on building three-dimensional dense maps. This paper focuses on representative systems of various V-SLAMs and gives their respective applicable scenarios and characteristics. Finally, this article forecasts the development of V-SLAM combining with multi-information fusion technology, semantic deep and learning technology.

Recently, there emerges a new kind of Web AR APP for brand advertising, which displays the brand-related content when a smartphone with the APP sees the brand's logo. However, the hand-crafted features used in these APPs usually lead to low recognition performance when used in a real environment. This is either because the target logo has less sharp edges and corner features which the hand-crafted schemes heavily depend on, or because some logos have larger intra-class variations and appear with different backgrounds. Some apps deal with this problem by just adding a button for the user to skip the phase of logo recognition. But this dramatically reduces the attractiveness of the apps as well as the advertisement effect. We promote an elegant solution to tackle the challenge of brand logo recognition. Thus, we propose a new CNN-based logo recognition scheme with a novel data enhancement strategy bias towards a single logo. We adapted the well-known CNN LeNet-5 to learn only two classes: the target logo and its best accompanying logo. Specifically, we utilized shape context as a measurement for characterizing the similarity between the target logo and other logos. Based on the similarity, the best accompanying dataset is generated for training the target logo. The experimental results show that the recall of our scheme surpasses hand-crafted feature based methods such as SIFT by 10% on real images while maintaining the same level of precision.

Logo recognition by Convolutional Neural Networks (CNNs) on a smartphone requires the network to be both accurate and small. In our previous work [1], we proposed the accompanying dataset method for single logo recognition to increase the recall and precision of the target logo recognition. However, the reason why it works was unclear, thus it was hard to compress the network while maintaining the same accuracy. In this paper, we use DeconvNet [9] to visualize our network’s feature maps and propose a metric to analyze them quantitatively. Finally, we obtain a better understanding of the influences in the network brought by accompanying datasets. Under its guidance, an effective way to compress the network is devised by us. The experiments show that we can reduce the size of the neural network’s first layer by 30% while only lower the recall and precision by 0.014 and 0.01. The training time is also saved by 40% due to the network compression.