• ETRI Journal
• /
• v.40 no.4
• /
• pp.421-434
• /
• 2018

The application of deep neural networks (DNNs) to connect the world with cyber physical systems (CPSs) has attracted much attention. However, DNNs require a large amount of memory and computational cost, which hinders their use in the relatively low-end smart devices that are widely used in CPSs. In this paper, we aim to determine whether DNNs can be efficiently deployed and operated in low-end smart devices. To do this, we develop a method to reduce the memory requirement of DNNs and increase the inference speed, while maintaining the performance (for example, accuracy) close to the original level. The parameters of DNNs are decomposed using a hybrid of canonical polyadic-singular value decomposition, approximated using a tensor power method, and fine-tuned by performing iterative one-shot hybrid fine-tuning to recover from a decreased accuracy. In this study, we evaluate our method on frequently used networks. We also present results from extensive experiments on the effects of several fine-tuning methods, the importance of iterative fine-tuning, and decomposition techniques. We demonstrate the effectiveness of the proposed method by deploying compressed networks in smartphones.

• Journal of Korea Multimedia Society
• /
• v.20 no.5
• /
• pp.769-781
• /
• 2017

Deep learning networks like Convolutional Neural Networks (CNNs) show successful performances in many computer vision applications such as image classification, object detection, and so on. For implementation of deep learning networks in embedded system with limited processing power and memory, deep learning network may need to be simplified. However, simplified deep learning network cannot learn every possible scene. One realistic strategy for embedded deep learning network is to construct a simplified deep learning network model optimized for the scene images of the installation place. Then, automatic training will be necessitated for commercialization. In this paper, as an intermediate step toward automatic training under fisheye camera environments, we study more precise human localization in fisheye images, and propose an accurate human localization method, Automatic Ground-Truth Labelling Method (AGTLM). AGTLM first localizes candidate human object bounding boxes by utilizing GoogLeNet-LSTM approach, and after reassurance process by GoogLeNet-based CNN network, finally refines them more correctly and precisely(tightly) by applying saliency object detection technique. The performance improvement of the proposed human localization method, AGTLM with respect to accuracy and tightness is shown through several experiments.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.6
• /
• pp.2748-2760
• /
• 2018

This paper proposes an audio event classification method based on convolutional neural networks (CNNs). CNN has great advantages of distinguishing complex shapes of image. Proposed system uses the features of audio sound as an input image of CNN. Mel scale filter bank features are extracted from each frame, then the features are concatenated over 40 consecutive frames and as a result, the concatenated frames are regarded as an input image. The output layer of CNN generates probabilities of audio event (e.g. dogs bark, siren, forest). The event probabilities for all images in an audio segment are accumulated, then the audio event having the highest accumulated probability is determined to be the classification result. This proposed method classified thirty audio events with the accuracy of 81.5% for the UrbanSound8K, BBC Sound FX, DCASE2016, and FREESOUND dataset.

• Journal of Korea Multimedia Society
• /
• v.22 no.11
• /
• pp.1269-1279
• /
• 2019

In medical image analysis, segmentation is considered as a vital process since it partitions an image into coherent parts and extracts interesting objects from the image. In this paper, we consider automatic segmentations of OCT retinal images to find six layer boundaries using convolutional neural networks. Segmenting retinal images by layer boundaries is very important in diagnosing and predicting progress of eye diseases including diabetic retinopathy, glaucoma, and AMD (age-related macular degeneration). We applied well-known CNN architecture for general image segmentation, called Segnet, U-net, and CNN-S into this problem. We also proposed a shortest path-based algorithm for finding the layer boundaries from the outputs of Segnet and U-net. We analysed their performance on public OCT image data set. The experimental results show that the Segnet combined with the proposed shortest path-based boundary finding algorithm outperforms other two networks.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.67 no.9
• /
• pp.1224-1231
• /
• 2018

Automatic age estimation has been used in many social network applications, practical commercial applications, and human-computer interaction visual-surveillance biometrics. However, it has rarely been explored. In this paper, we propose an automatic age estimation system, which includes face detection and convolutional deep learning based on an inception module. The latter is a 22-layer-deep network that serves as the particular category of the inception design. To evaluate the proposed approach, we use 4,000 images of eight different age groups from the Adience age dataset. k-fold cross-validation (k = 5) is applied. A comparison of the performance of the proposed work and recent related methods is presented. The results show that the proposed method significantly outperforms existing methods in terms of the exact accuracy and off-by-one accuracy. The off-by-one accuracy is when the result is off by one adjacent age label to the above or below. For the exact accuracy, the age label of "60+" is classified with the highest accuracy of 76%.

• KIPS Transactions on Software and Data Engineering
• /
• v.5 no.11
• /
• pp.607-614
• /
• 2016

In this paper, we propose an intelligent fire learning and detection system using convolutional neural networks (CNN). Through the convolutional layer of the CNN, various features of flame and smoke images are automatically extracted, and these extracted features are learned to classify them into flame or smoke or no fire. In order to detect fire in the image, candidate fire regions are first extracted from the image and extracted candidate regions are passed through CNN. Experimental results on various image shows that our system has better performances over previous work.

• Journal of the Korean Wood Science and Technology
• /
• v.49 no.5
• /
• pp.491-503
• /
• 2021

Wood identification is regularly performed by observing the wood anatomy, such as colour, texture, fibre direction, and other characteristics. The manual process, however, could be time consuming, especially when identification work is required at high quantity. Considering this condition, a convolutional neural networks (CNN)-based program is applied to improve the image classification results. The research focuses on the algorithm accuracy and efficiency in dealing with the dataset limitations. For this, it is proposed to do the sample selection process or only take a small portion of the existing image. Still, it can be expected to represent the overall picture to maintain and improve the generalisation capabilities of the CNN method in the classification stages. The experiments yielded an incredible F1 score average up to 93.4% for medium sample area sizes (200 × 200 pixels) on each CNN architecture (VGG16, ResNet50, MobileNet, DenseNet121, and Xception based). Whereas DenseNet121-based architecture was found to be the best architecture in maintaining the generalisation of its model for each sample area size (100, 200, and 300 pixels). The experimental results showed that the proposed algorithm can be an accurate and reliable solution.

• Journal of the Korean Wood Science and Technology
• /
• v.47 no.4
• /
• pp.385-392
• /
• 2019

In our previous study, we investigated the use of ensemble models based on LeNet and MiniVGGNet to classify the images of transverse and longitudinal surfaces of five Korean softwoods (cedar, cypress, Korean pine, Korean red pine, and larch). It had accomplished an average F1 score of more than 98%; the classification performance of the longitudinal surface image was still less than that of the transverse surface image. In this study, ensemble methods of two different convolutional neural network models (LeNet3 for smartphone camera images and NIRNet for NIR spectra) were applied to lumber species classification. Experimentally, the best classification performance was obtained by the averaging ensemble method of LeNet3 and NIRNet. The average F1 scores of the individual LeNet3 model and the individual NIRNet model were 91.98% and 85.94%, respectively. By the averaging ensemble method of LeNet3 and NIRNet, an average F1 score was increased to 95.31%.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.2
• /
• pp.130-138
• /
• 2021

This paper presents an effective structure by applying various normalization to Convolutional Neural Networks (CNN) for seismic event classification. Normalization techniques can not only improve the learning speed of neural networks, but also show robustness to noise. In this paper, we analyze the effect of input data normalization and hidden layer normalization on the deep learning model for seismic event classification. In addition an effective model is derived through various experiments according to the structure of the applied hidden layer. As a result of various experiments, the model that applied input data normalization and weight normalization to the first hidden layer showed the most stable performance improvement.

• Structural Engineering and Mechanics
• /
• v.77 no.4
• /
• pp.495-508
• /
• 2021

This study devises a novel approach, namely quadruple 1D convolutional neural network, for detecting connection stiffness reduction in steel truss bridge structure using experimental and numerical modal data. The method is developed based on expertise in two domains: firstly, in Structural Health Monitoring, the mode shapes and its high-order derivatives, including second, third, and fourth derivatives, are accurate indicators in assessing damages. Secondly, in the Machine Learning literature, the deep convolutional neural networks are able to extract relevant features from input data, then perform classification tasks with high accuracy and reduced time complexity. The efficacy and effectiveness of the present method are supported through an extensive case study with the railway Nam O bridge. It delivers highly accurate results in assessing damage localization and damage severity for single as well as multiple damage scenarios. In addition, the robustness of this method is tested with the presence of white noise reflecting unavoidable uncertainties in signal processing and modeling in reality. The proposed approach is able to provide stable results with data corrupted by noise up to 10%.