• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.553-556
• /
• 2020

최근 이미지 인식을 위한 Convolutional Neural Network(CNN) 모델의 경량화에 관한 연구가 활발하게 이루어지고 있다. 그중 양자화는 모델을 구성하는 가중치의 크기를 낮추는 방법이다. 기존의 CNN 모델에서 가장 큰 비중을 하는 Fully Connected Layer(FCL)는 내부적으로 32 Bit의 실수 행렬로 표현된다. 본 논문에서는 미리 학습된 실수 가중치를 더 작은 비트의 정수 행렬로 양자화한다. 양자화된 행렬에 대해서 영상 압축 등에서 사용하는 Discrete Cosine Transform(DCT)을 통해 주파수 영역으로 변환한 후 고주파 영역을 생략하는 손실압축 방법을 제안한다. 실험을 통해 그 과정에서 손실에 따른 정확도의 변화를 나타낸다.

• Journal of Intelligence and Information Systems
• /
• v.24 no.1
• /
• pp.205-225
• /
• 2018

Convolutional Neural Network (ConvNet) is one class of the powerful Deep Neural Network that can analyze and learn hierarchies of visual features. Originally, first neural network (Neocognitron) was introduced in the 80s. At that time, the neural network was not broadly used in both industry and academic field by cause of large-scale dataset shortage and low computational power. However, after a few decades later in 2012, Krizhevsky made a breakthrough on ILSVRC-12 visual recognition competition using Convolutional Neural Network. That breakthrough revived people interest in the neural network. The success of Convolutional Neural Network is achieved with two main factors. First of them is the emergence of advanced hardware (GPUs) for sufficient parallel computation. Second is the availability of large-scale datasets such as ImageNet (ILSVRC) dataset for training. Unfortunately, many new domains are bottlenecked by these factors. For most domains, it is difficult and requires lots of effort to gather large-scale dataset to train a ConvNet. Moreover, even if we have a large-scale dataset, training ConvNet from scratch is required expensive resource and time-consuming. These two obstacles can be solved by using transfer learning. Transfer learning is a method for transferring the knowledge from a source domain to new domain. There are two major Transfer learning cases. First one is ConvNet as fixed feature extractor, and the second one is Fine-tune the ConvNet on a new dataset. In the first case, using pre-trained ConvNet (such as on ImageNet) to compute feed-forward activations of the image into the ConvNet and extract activation features from specific layers. In the second case, replacing and retraining the ConvNet classifier on the new dataset, then fine-tune the weights of the pre-trained network with the backpropagation. In this paper, we focus on using multiple ConvNet layers as a fixed feature extractor only. However, applying features with high dimensional complexity that is directly extracted from multiple ConvNet layers is still a challenging problem. We observe that features extracted from multiple ConvNet layers address the different characteristics of the image which means better representation could be obtained by finding the optimal combination of multiple ConvNet layers. Based on that observation, we propose to employ multiple ConvNet layer representations for transfer learning instead of a single ConvNet layer representation. Overall, our primary pipeline has three steps. Firstly, images from target task are given as input to ConvNet, then that image will be feed-forwarded into pre-trained AlexNet, and the activation features from three fully connected convolutional layers are extracted. Secondly, activation features of three ConvNet layers are concatenated to obtain multiple ConvNet layers representation because it will gain more information about an image. When three fully connected layer features concatenated, the occurring image representation would have 9192 (4096+4096+1000) dimension features. However, features extracted from multiple ConvNet layers are redundant and noisy since they are extracted from the same ConvNet. Thus, a third step, we will use Principal Component Analysis (PCA) to select salient features before the training phase. When salient features are obtained, the classifier can classify image more accurately, and the performance of transfer learning can be improved. To evaluate proposed method, experiments are conducted in three standard datasets (Caltech-256, VOC07, and SUN397) to compare multiple ConvNet layer representations against single ConvNet layer representation by using PCA for feature selection and dimension reduction. Our experiments demonstrated the importance of feature selection for multiple ConvNet layer representation. Moreover, our proposed approach achieved 75.6% accuracy compared to 73.9% accuracy achieved by FC7 layer on the Caltech-256 dataset, 73.1% accuracy compared to 69.2% accuracy achieved by FC8 layer on the VOC07 dataset, 52.2% accuracy compared to 48.7% accuracy achieved by FC7 layer on the SUN397 dataset. We also showed that our proposed approach achieved superior performance, 2.8%, 2.1% and 3.1% accuracy improvement on Caltech-256, VOC07, and SUN397 dataset respectively compare to existing work.

• International journal of advanced smart convergence
• /
• v.10 no.1
• /
• pp.105-116
• /
• 2021

In order to preserve the seeds of excellent Hanwoo(Korean traditional cattle) and secure quality competitiveness in the infinite competition with foreign imported beef, production of high-quality Hanwoo beef is absolutely necessary. %IMF (Intramuscular Fat Percentage) is one of the most important factors in evaluating the value of high-quality meat, although standards vary according to food culture and industrial conditions by country. Therefore, it is required to develop a %IMF estimation algorithm suitable for Hanwoo. In this study, we proposed a method of estimating %IMF of Hanwoo using CNN in ultrasound images. First, the proposed method classified the chemically measured %IMF into 10 classes using k-means clustering method to apply CNN. Next, ROI images were obtained at regular intervals from each ultrasound image and used for CNN training and estimation. The proposed CNN model is composed of three stages of convolution layer and fully connected layer. As a result of the experiment, it was confirmed that the %IMF of Hanwoo was estimated with an accuracy of 98.2%. The correlation coefficient between the estimated %IMF and the real %IMF by the proposed method is 0.97, which is about 10% better than the 0.88 of the previous method.

• Journal of the Computational Structural Engineering Institute of Korea
• /
• v.36 no.4
• /
• pp.221-232
• /
• 2023

This paper presents a novel convolutional neural network (CNN)-based approach for predicting bolt clamping force in the early bolt loosening state of bolted structures. The approach entails tightening eight bolts with different clamping forces and generating frequency responses, which are then used to create a similarity map. This map quantifies the magnitude and shape similarity between the frequency responses and the initial model in a fully fastened state. Krylov subspace-based model order reduction is employed to efficiently handle the large amount of frequency response data. The CNN model incorporates a regression output layer to predict the clamping forces of the bolts. Its performance is evaluated by training the network by using various amounts of training data and convolutional layers. The input data for the model are derived from the magnitude and shape similarity map obtained from the frequency responses. The results demonstrate the diagnostic potential and effectiveness of the proposed approach in detecting early bolt loosening. Accurate bolt clamping force predictions in the early loosening state can thus be achieved by utilizing the frequency response data and CNN model. The findings afford valuable insights into the application of CNNs for assessing the integrity of bolted structures.

• Nuclear Engineering and Technology
• /
• v.54 no.6
• /
• pp.2297-2310
• /
• 2022

A condensation heat transfer model is essential to accurately predict the performance of the passive containment cooling system (PCCS) during an accident in an advanced light water reactor. However, most of existing models tend to predict condensation heat transfer very well for a specific range of thermal-hydraulic conditions. In this study, a new correlation for condensation heat transfer coefficient (HTC) is presented using machine learning technique. To secure sufficient training data, a large number of pseudo data were produced by using ten existing condensation models. Then, a neural network model was developed, consisting of a fully connected layer and a convolutional neural network (CNN) algorithm, DenseNet. Based on the hold-out cross-validation, the neural network was trained and validated against the pseudo data. Thereafter, it was evaluated using the experimental data, which were not used for training. The machine learning model predicted better results than the existing models. It was also confirmed through a parametric study that the machine learning model presents continuous and physical HTCs for various thermal-hydraulic conditions. By reflecting the effects of individual variables obtained from the parametric analysis, a new correlation was proposed. It yielded better results for almost all experimental conditions than the ten existing models.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.29-32
• /
• 2020

Due to the lack of improper image acquisition process, noise induction is an inevitable step. As a result, objective image quality assessment (IQA) plays an important role in estimating the visual quality of noisy image. Plenty of IQA methods have been proposed including traditional signal processing based methods as well as current deep learning based methods where the later one shows promising performance due to their complex representation ability. The deep learning based methods consists of several convolution layers and down sampling layers for feature extraction and fully connected layers for regression. Usually, the down sampling is performed by using max-pooling layer after each convolutional block. We reveal that this max-pooling causes information loss despite of knowing their importance. Consequently, we propose a better IQA method that replaces the max-pooling layers with strided convolutions to down sample the feature space and since the strided convolution layers have learnable parameters, they preserve optimal features and discard redundant information, thereby improve the prediction accuracy. The experimental results verify the effectiveness of the proposed method.

• Journal of Information Processing Systems
• /
• v.19 no.5
• /
• pp.673-687
• /
• 2023

Convolutional neural network-based deep learning technology is the most commonly used in image identification, but it requires large-scale data for training. Therefore, application in specific fields in which data acquisition is limited, such as in the military, may be challenging. In particular, the identification of ground weapon systems is a very important mission, and high identification accuracy is required. Accordingly, various studies have been conducted to achieve high performance using small-scale data. Among them, the ensemble method, which achieves excellent performance through the prediction average of the pre-trained models, is the most representative method; however, it requires considerable time and effort to find the optimal combination of ensemble models. In addition, there is a performance limitation in the prediction results obtained by using an ensemble method. Furthermore, it is difficult to obtain the ensemble effect using models with imbalanced classification accuracies. In this paper, we propose a transfer learning-based feature fusion technique for heterogeneous models that extracts and fuses features of pre-trained heterogeneous models and finally, fine-tunes hyperparameters of the fully connected layer to improve the classification accuracy. The experimental results of this study indicate that it is possible to overcome the limitations of the existing ensemble methods by improving the classification accuracy through feature fusion between heterogeneous models based on transfer learning.

• The Journal of the Acoustical Society of Korea
• /
• v.36 no.6
• /
• pp.407-412
• /
• 2017

In this paper, we propose a speech emotion recognition method using a deep neural network based on the attention mechanism. The proposed method consists of a combination of CNN (Convolution Neural Networks), GRU (Gated Recurrent Unit), DNN (Deep Neural Networks) and attention mechanism. The spectrogram of the speech signal contains characteristic patterns according to the emotion. Therefore, we modeled characteristic patterns according to the emotion by applying the tuned Gabor filters as convolutional filter of typical CNN. In addition, we applied the attention mechanism with CNN and FC (Fully-Connected) layer to obtain the attention weight by considering context information of extracted features and used it for emotion recognition. To verify the proposed method, we conducted emotion recognition experiments on six emotions. The experimental results show that the proposed method achieves higher performance in speech emotion recognition than the conventional methods.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.16 no.1
• /
• pp.41-49
• /
• 2023

A convolutional neural network (CNN) is a representative algorithm for implementing artificial neural networks. CNNs have improved on the issues of rapid increase in calculation amount and low object classification rates, which are associated with a conventional multi-layered fully-connected neural network (FNN). However, because of the rapid development of IT devices, the maximum resolution of images captured by current smartphone and tablet cameras has reached 108 million pixels (MP). Specifically, a traditional CNN algorithm requires a significant cost and time to learn and process simple, high-resolution images. Therefore, this study proposes an improved CNN algorithm for implementing an object classification learning model for simple, high-resolution images. The proposed method alters the adjacency matrix value of the pooling layer's max pooling operation for the CNN algorithm to reduce the high-resolution image learning model's creation time. This study implemented a learning model capable of processing 4, 8, and 12 MP high-resolution images for each altered matrix value. The performance evaluation result showed that the creation time of the learning model implemented with the proposed algorithm decreased by 36.26% for 12 MP images. Compared to the conventional model, the proposed learning model's object recognition accuracy and loss rate were less than 1%, which is within the acceptable error range. Practical verification is necessary through future studies by implementing a learning model with more varied image types and a larger amount of image data than those used in this study.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.10
• /
• pp.3729-3749
• /
• 2021

At present, deep convolution network-based salient object detection (SOD) has achieved impressive performance. However, it is still a challenging problem to make full use of the multi-scale information of the extracted features and which appropriate feature fusion method is adopted to process feature mapping. In this paper, we propose a new adjacency auxiliary network (AANet) based on multi-scale feature fusion for SOD. Firstly, we design the parallel connection feature enhancement module (PFEM) for each layer of feature extraction, which improves the feature density by connecting different dilated convolution branches in parallel, and add channel attention flow to fully extract the context information of features. Then the adjacent layer features with close degree of abstraction but different characteristic properties are fused through the adjacent auxiliary module (AAM) to eliminate the ambiguity and noise of the features. Besides, in order to refine the features effectively to get more accurate object boundaries, we design adjacency decoder (AAM_D) based on adjacency auxiliary module (AAM), which concatenates the features of adjacent layers, extracts their spatial attention, and then combines them with the output of AAM. The outputs of AAM_D features with semantic information and spatial detail obtained from each feature are used as salient prediction maps for multi-level feature joint supervising. Experiment results on six benchmark SOD datasets demonstrate that the proposed method outperforms similar previous methods.