• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.10
• /
• pp.3668-3684
• /
• 2021

Video action recognition is widely used in video surveillance, behavior detection, human-computer interaction, medically assisted diagnosis and motion analysis. However, video action recognition can be disturbed by many factors, such as background, illumination and so on. Two-stream convolutional neural network uses the video spatial and temporal models to train separately, and performs fusion at the output end. The multi segment Two-Stream convolutional neural network model trains temporal and spatial information from the video to extract their feature and fuse them, then determine the category of video action. Google Xception model and the transfer learning is adopted in this paper, and the Xception model which trained on ImageNet is used as the initial weight. It greatly overcomes the problem of model underfitting caused by insufficient video behavior dataset, and it can effectively reduce the influence of various factors in the video. This way also greatly improves the accuracy and reduces the training time. What's more, to make up for the shortage of dataset, the kinetics400 dataset was used for pre-training, which greatly improved the accuracy of the model. In this applied research, through continuous efforts, the expected goal is basically achieved, and according to the study and research, the design of the original dual-flow model is improved.

• Journal of Information Processing Systems
• /
• v.17 no.2
• /
• pp.399-410
• /
• 2021

Aiming at the problem that complex external variables in natural scenes have a greater impact on facial expression recognition results, a facial expression recognition method based on two-stream convolutional neural network is proposed. The model introduces exponentially enhanced shared input weights before each level of convolution input, and uses soft attention mechanism modules on the space-time features of the combination of static and dynamic streams. This enables the network to autonomously find areas that are more relevant to the expression category and pay more attention to these areas. Through these means, the information of irrelevant interference areas is suppressed. In order to solve the problem of poor local robustness caused by lighting and expression changes, this paper also performs lighting preprocessing with the lighting preprocessing chain algorithm to eliminate most of the lighting effects. Experimental results on AFEW6.0 and Multi-PIE datasets show that the recognition rates of this method are 95.05% and 61.40%, respectively, which are better than other comparison methods.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.5
• /
• pp.2253-2272
• /
• 2018

We propose a deep learning method for multi-focus image fusion. Unlike most existing pixel-level fusion methods, either in spatial domain or in transform domain, our method directly learns an end-to-end fully convolutional two-stream network. The framework maps a pair of different focus images to a clean version, with a chain of convolutional layers, fusion layer and deconvolutional layers. Our deep fusion model has advantages of efficiency and robustness, yet demonstrates state-of-art fusion quality. We explore different parameter settings to achieve trade-offs between performance and speed. Moreover, the experiment results on our training dataset show that our network can achieve good performance with subjective visual perception and objective assessment metrics.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2021.11a
• /
• pp.878-879
• /
• 2021

Most state-of-the-art CNNs for action recognition are based on a two-stream architecture: RGB frames stream represents the appearance and the optical flow stream interprets the motion of action. However, the cost of optical flow computation is very high and then it increases action recognition latency. We introduce a design strategy for action recognition inspired by a two-stream network and teacher-student architecture. There are two sub-networks in our neural networks, the optical flow sub-network as a teacher and the RGB frames sub-network as a student. In the training stage, we distill the feature from the teacher as a baseline to train student sub-network. In the test stage, we only use the student so that the latency reduces without computing optical flow. Our experiments show that its advantages over two-stream architecture in both speed and performance.

• The Journal of Korean Institute of Next Generation Computing
• /
• v.14 no.6
• /
• pp.66-74
• /
• 2018

Recently, there has been active studies on hand gesture recognition to increase immersion and provide user-friendly interaction in a virtual reality environment. However, most studies require specialized sensors or equipment, or show low recognition rates. This paper proposes a hand gesture recognition method using Deep Learning technology without separate sensors or equipment other than camera to recognize static and dynamic hand gestures. First, a series of hand gesture input images are converted into high-frequency images, then each of the hand gestures RGB images and their high-frequency images is learned through the DenseNet three-dimensional Convolutional Neural Network. Experimental results on 6 static hand gestures and 9 dynamic hand gestures showed an average of 92.6% recognition rate and increased 4.6% compared to previous DenseNet. The 3D defense game was implemented to verify the results of our study, and an average speed of 30 ms of gesture recognition was found to be available as a real-time user interface for virtual reality applications.

• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.3
• /
• pp.99-104
• /
• 2022

Convolutional neural network (CNN) generally uses two-stream architecture RGB and optical flow stream for its action recognition function. RGB frames stream display appearance and optical flow stream interprets its action. However, the standard method of using optical flow is costly in its computational time and latency associated with increased action recognition. The purpose of the study was to evaluate a novel way to create a two sub-networks in neural networks. The optical flow sub-network was assigned as a teacher and the RGB frames as a student. In the training stage, the optical flow sub-network extracts features through the teacher sub-network and transmits the information to student sub-network for baseline training. In the test stage, only student sub-network was operational with decreased in latency without computing optical flow. Experimental results shows that our network fed only by RGB stream gets a competitive accuracy of 54.5% on HMDB51, which is 1.5 times better than that on R3D-18.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.7
• /
• pp.3599-3619
• /
• 2019

In human activity recognition system both static and motion information play crucial role for efficient and competitive results. Most of the existing methods are insufficient to extract video features and unable to investigate the level of contribution of both (Static and Motion) components. Our work highlights this problem and proposes Static-Motion fused features descriptor (SMFD), which intelligently leverages both static and motion features in the form of descriptor. First, static features are learned by two-stream 3D convolutional neural network. Second, trajectories are extracted by tracking key points and only those trajectories have been selected which are located in central region of the original video frame in order to to reduce irrelevant background trajectories as well computational complexity. Then, shape and motion descriptors are obtained along with key points by using SIFT flow. Next, cholesky transformation is introduced to fuse static and motion feature vectors to guarantee the equal contribution of all descriptors. Finally, Long Short-Term Memory (LSTM) network is utilized to discover long-term temporal dependencies and final prediction. To confirm the effectiveness of the proposed approach, extensive experiments have been conducted on three well-known datasets i.e. UCF101, HMDB51 and YouTube. Findings shows that the resulting recognition system is on par with state-of-the-art methods.