• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.06a
• /
• pp.98-101
• /
• 2019

At present, deep convolutional neural networks have made a very important contribution in single-image super-resolution. Through the learning of the neural networks, the features of input images are transformed and combined to establish a nonlinear mapping of low-resolution images to high-resolution images. Some previous methods are difficult to train and take up a lot of memory. In this paper, we proposed a simple and compact deep recursive residual network learning the features for single image super resolution. Global residual learning and local residual learning are used to reduce the problems of training deep neural networks. And the recursive structure controls the number of parameters to save memory. Experimental results show that the proposed method improved image qualities that occur in previous methods.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.11
• /
• pp.4557-4572
• /
• 2020

The Alaska competition provides an opportunity to study the practical problems of real-world steganalysis. Participants are required to solve steganalysis involving various embedding schemes, inconsistency JPEG Quality Factor and various processing pipelines. In this paper, we propose a method to ensemble multiple deep learning steganalyzers. We select SRNet and RESDET as our base models. Then we design a three-layers model ensemble network to fuse these base models and output the final prediction. By separating the three colors channels for base model training and feature replacement strategy instead of simply merging features, the performance of the model ensemble is greatly improved. The proposed method won second place in the Alaska 1 competition in the end.

• The Journal of the Korea Contents Association
• /
• v.6 no.3
• /
• pp.12-19
• /
• 2006

In this paper, we describe a deep analysis of question for question answering system. It is difficult to offer the correct answer because general question answering systems do not analyze the semantic of user's natural language question. We analyze user's question semantically and extract semantic features using the semantic feature extraction grammar and characteristics of natural language question. They are represented as semantic features and grammatical morphemes that consider semantic and syntactic structure of user's questions. We evaluated our approach using 100 questions whose answer type is a person in the web. We showed that a deep analysis of questions which are comparatively short but enough to mean can analysis the user's intention and extract semantic features.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.9
• /
• pp.2483-2504
• /
• 2023

Most existing low-light enhancement algorithms either use a large number of training parameters or lack generalization to real-world scenarios. This paper presents a novel lightweight and robust pixel-wise polynomial approximation-based deep network for low-light image enhancement. For mapping the low-light image to the enhanced image, pixel-wise higher-order polynomials are employed. A deep convolution network is used to estimate the coefficients of these higher-order polynomials. The proposed network uses multiple branches to estimate pixel values based on different receptive fields. With a smaller receptive field, the first branch enhanced local features, the second and third branches focused on medium-level features, and the last branch enhanced global features. The low-light image is downsampled by the factor of 2b-1 (b is the branch number) and fed as input to each branch. After combining the outputs of each branch, the final enhanced image is obtained. A comprehensive evaluation of our proposed network on six publicly available no-reference test datasets shows that it outperforms state-of-the-art methods on both quantitative and qualitative measures.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.9
• /
• pp.2361-2376
• /
• 2023

The electrocardiogram (ECG) signal is commonly used to screen and diagnose cardiovascular diseases. In recent years, deep neural networks have been regarded as an effective way for automatic ECG disease diagnosis. The convolutional neural network is widely used for ECG signal extraction because it can obtain different levels of information. However, most previous studies adopt single scale convolution filters to extract ECG signal features, ignoring the complementarity between ECG signal features of different scales. In the paper, we propose a dual-scale network with convolution filters of different sizes for 12-lead ECG classification. Our model can extract and fuse ECG signal features of different scales. In addition, different spatial and time periods of the feature map obtained from the 12-lead ECG may have different contributions to ECG classification. Therefore, we add a spatial-temporal attention to each scale sub-network to emphasize the representative local spatial and temporal features. Our approach is evaluated on PTB-XL dataset and achieves 0.9307, 0.8152, and 89.11 on macro-averaged ROC-AUC score, a maximum F1 score, and mean accuracy, respectively. The experiment results have proven that our approach outperforms the baselines.

• Journal of Practical Engineering Education
• /
• v.16 no.5_spc
• /
• pp.611-618
• /
• 2024

The rapid advancement of deep learning has significantly enhanced the performance of single image super-resolution (SR). However, most existing deep learning-based image SR networks only facilitate information flow in the forward direction, which limits their performance. In this study, we investigate a feedback network for precise image SR. This feedback network effectively enhances lower-level feature representation by rerouting multiple higher-level features. We sequentially construct several Residual Density Modules and deploy them repeatedly over time. Multiple feedback connections between two adjacent time steps leverage high-level features captured within a large receptive field to refine low-level features lacking sufficient contextual information. A carefully designed feedback module efficiently selects and enhances valuable information from the rerouted high-level features, thereby improving low-level features with enriched high-level information. Extensive experiments demonstrate that the proposed method outperforms existing approaches in both objective and subjective evaluations.

• ETRI Journal
• /
• v.39 no.6
• /
• pp.832-840
• /
• 2017

Recently, deep recurrent neural networks have achieved great success in various machine learning tasks, and have also been applied for sound event detection. The detection of temporally overlapping sound events in realistic environments is much more challenging than in monophonic detection problems. In this paper, we present an approach to improve the accuracy of polyphonic sound event detection in multichannel audio based on gated recurrent neural networks in combination with auditory spectral features. In the proposed method, human hearing perception-based spatial and spectral-domain noise-reduced harmonic features are extracted from multichannel audio and used as high-resolution spectral inputs to train gated recurrent neural networks. This provides a fast and stable convergence rate compared to long short-term memory recurrent neural networks. Our evaluation reveals that the proposed method outperforms the conventional approaches.

• ETRI Journal
• /
• v.41 no.6
• /
• pp.750-759
• /
• 2019

Presently, the extraction of hand-crafted features is still the dominant method in radar emitter recognition. To solve the complicated problems of selection and updation of empirical features, we present a novel automatic feature extraction structure based on deep learning. In particular, a convolutional neural network (CNN) is adopted to extract high-level abstract representations from the time-frequency images of emitter signals. Thus, the redundant process of designing discriminative features can be avoided. Furthermore, to address the performance degradation of a single platform, we propose the construction of an ensemble learning-based architecture for multi-platform fusion recognition. Experimental results indicate that the proposed algorithms are feasible and effective, and they outperform other typical feature extraction and fusion recognition methods in terms of accuracy. Moreover, the proposed structure could be extended to other prevalent ensemble learning alternatives.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.2
• /
• pp.832-854
• /
• 2019

Action recognition has been studied in computer vision field for years. We present an effective approach to recognize actions using a dimension reduction method, which is applied as a crucial step to reduce the dimensionality of feature descriptors after extracting features. We propose to use sparse matrix and randomized kd-tree to modify it and then propose modified Local Fisher Discriminant Analysis (mLFDA) method which greatly reduces the required memory and accelerate the standard Local Fisher Discriminant Analysis. For feature encoding, we propose a useful encoding method called mix encoding which combines Fisher vector encoding and locality-constrained linear coding to get the final video representations. In order to add more meaningful features to the process of action recognition, the convolutional neural network is utilized and combined with mix encoding to produce the deep network feature. Experimental results show that our algorithm is a competitive method on KTH dataset, HMDB51 dataset and UCF101 dataset when combining all these methods.

• International Journal of Computer Science & Network Security
• /
• v.23 no.8
• /
• pp.40-48
• /
• 2023

Speech Emotions recognition has become the active research theme in speech processing and in applications based on human-machine interaction. In this work, our system is a two-stage approach, namely feature extraction and classification engine. Firstly, two sets of feature are investigated which are: the first one is extracting only 13 Mel-frequency Cepstral Coefficient (MFCC) from emotional speech samples and the second one is applying features fusions between the three features: Zero Crossing Rate (ZCR), Teager Energy Operator (TEO), and Harmonic to Noise Rate (HNR) and MFCC features. Secondly, we use two types of classification techniques which are: the Support Vector Machines (SVM) and the k-Nearest Neighbor (k-NN) to show the performance between them. Besides that, we investigate the importance of the recent advances in machine learning including the deep kernel learning. A large set of experiments are conducted on Surrey Audio-Visual Expressed Emotion (SAVEE) dataset for seven emotions. The results of our experiments showed given good accuracy compared with the previous studies.