Search

• We propose a unified algorithm to compress the CNN for IAHCCR. • We achieve 12.4 × compression and 1.7 × acceleration with only 0.17% accuracy loss. • The experiments are conducted on IAHCC-UCAS2016, ICDAR-2013, and MNIST. • The state-of-the-art performance is achieved on IAHCC-UCAS2016. Since the convolutional neural network (CNN) has brought great breakthroughs in the field of computer vision, it recently has been introduced to the in-air handwritten Chinese character recognition (IAHCCR) to achieve better recognition performance. However, the CNN is typically over-parameterized and contains lots of redundant filters or parameters. This leads the CNN to suffer from huge computation cost and considerable storage usage, limiting its deployments to resource-constrained devices like mobile phones and intelligent TVs. In this paper, we propose a unified algorithm to effectively compress the CNN for IAHCCR with little accuracy loss. Specifically, we first utilize the channel pruning strategy to simplify the network structure, and then adopt the network quantization technique to represent parameters with lower precision. We conduct experiments on the in-air handwriting dataset IAHCC-UCAS2016, where the baseline CNN achieves the state-of-the-art accuracy of 95.33% with 15.5 MB of storage. After the compression, we achieve 12.4 × storage saving and 1.7 × theoretical acceleration with only 0.17% accuracy loss. Moreover, evaluations on other benchmark datasets including ICDAR-2013 and MNIST further demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Highlights • The paper is theoretically well founded and motivated by recent research on the physiological aspects of visual recognition. • The proposed idea is simple yet effective. It provides a practical way of training CNN for cross-domain 3D shape retrieval. • High versatility, a single framework for all the three investigated tasks. • A new selected-view matching strategy based on-line learning. • The experiments are sufficient with significantly improved results. Graphical abstract Abstract Recent vision-brain physiological experiments [65] have demonstrated that contours, and in particular, contour junctions present in the 2D images are very informative for revealing the 3D structure of the object. Inspired by this observation, we take 2D sketches (or 2D views of 3D sketches) and edge maps of 2D views of 3D models as a unified domain to train the Convolutional Neural Network (CNN). The CNN features are then used for 3D object representation. We show that the CNN can successfully learn the object structure from different types of clues. The performance of the proposed method demonstrates that the semantic gap between the 2D/3D sketches and the 3D models can be bridged without any cross-domain similarity learning. Experiments show that our approach significantly outperforms the state-of-the-art 2D/3D sketch-based 3D retrieval methods. [ABSTRACT FROM AUTHOR]

Human action recognition in video analytics has been widely studied in recent years. Yet, most of these methods assign a single action label to video after either analyzing a complete video or using classifier for each frame. But when compared to human vision strategy, it can be deduced that we (human) require just an instance of visual data for recognition of scene. It turns out that small group of frames or even single frame from the video are enough for precise recognition. In this paper, we present an approach to detect, localize and recognize actions of interest in almost real-time from frames obtained by a continuous stream of video data that can be captured from a surveillance camera. The model takes input frames after a specified period and is able to give action label based on a single frame. Combining results over specific time we predicted the action label for the stream of video. We demonstrate that YOLO is effective method and comparatively fast for recognition and localization in Liris Human Activities dataset. [ABSTRACT FROM AUTHOR]

Like other problems in computer vision, offline handwritten Chinese character recognition (HCCR) has achieved impressive results using convolutional neural network (CNN)-based methods. However, larger and deeper networks are needed to deliver state-of-the-art results in this domain. Such networks intuitively appear to incur high computational cost, and require the storage of a large number of parameters, which render them unfeasible for deployment in portable devices. To solve this problem, we propose a Global Supervised Low-rank Expansion (GSLRE) method and an Adaptive Drop-weight (ADW) technique to solve the problems of speed and storage capacity. We design a nine-layer CNN for HCCR consisting of 3755 classes, and devise an algorithm that can reduce the network’s computational cost by nine times and compress the network to 1/18 of the original size of the baseline model, with only a 0.21% drop in accuracy. In tests, the proposed algorithm can still surpass the best single-network performance reported thus far in the literature while requiring only 2.3MB for storage. Furthermore, when integrated with our effective forward implementation, the recognition of an offline character image takes only 9.7 ms on a CPU. Compared with the state-of-the-art CNN model for HCCR, our approach is approximately 30 times faster, yet 10 times more cost efficient. [ABSTRACT FROM AUTHOR]

Action recognition (AR) is one of the most important tasks in video analysis and computer vision. Recently a large number of related methods have been proposed. While most of these methods are investigated on AR datasets collected from the visible spectrum, the AR problem under infrared scenarios still has not attracted much attention. There is even few public infrared datasets available for supporting the fundamental evaluation requirements of this research. To this issue, this work aims to emphasize the importance of the infrared AR problem in applications and arouse researchers' attention on this task. Specifically, we construct a new Inf rared A ction R ecognition ( InfAR ) dataset captured at different times, including in summer and winter, and explore how discriminable actions in our InfAR dataset are with the state-of-the-art pipelines based on low-level features and deep convolutional neural network (CNN), respectively. Our results reveal: (1) In all, dense trajectory feature can achieve the best performance while the appearance features, e.g., HOG, have relatively poorer performance; (2) the encoding method of vector of locally aggregated descriptors is evidently better than that of the widely-used Fisher Vector; (3) the late fusion facilitates a better performance than early fusion; (4) action videos captured in winter is more discriminable than in summer; (5) compared to appearance information, the motion information is more essential for infrared action recognition and utilizing this information through deep CNN can improve greatly the performance. The best performance achieved on our dataset is 76.66% (Average Precision), leaving a reasonable space for further exploring the insights underlying such type of infrared AR problem and accordingly designing proper techniques to further promote the performance on this specifically constructed InfAR dataset. [ABSTRACT FROM AUTHOR]

Scene text detection and scene segmentation are meaningful tasks in the computer vision field. Could the semantic scene segmentation assist scene text detection in any degree? For example, can we expect the probability of a region being text is low if its surrounding segment, i.e., its context, is labeled as sky? In this paper, we have a positive answer by constructing a scene context-based text detection model. In this model, we use texton features and a fully-connected conditional random field (CRF) to estimate pixel-level scene class׳s probability to be considered as image׳s context feature. Meanwhile, maximally stable extremal regions (MSERs) are extracted, integrated and extended as image patches of character candidates. Then, each image patch is fed to a simple two-layer convolutional neural network (CNN) to automatically extract its character feature. The averaged context feature of the corresponding patch is considered as the patch׳s context feature. The character feature and context feature are fused as the input into a support vector machine for text/non-text determination. Finally, as a post-processing, neighboring text regions are grouped hierarchically. The performance evaluation on ICDAR2013 and SVT databases, as well as a preliminary evaluation on a patch-level database, proves that the scene context can improve the performance of scene text detection. Moreover, the comparative study with state-of-the-art methods shows the top-level performance of our method. [ABSTRACT FROM AUTHOR]