• Proceedings of the Korea Information Processing Society Conference
• /
• 2016.04a
• /
• pp.633-635
• /
• 2016

Convolutional Networks (ConvNets) are powerful models that learn hierarchies of visual features, which could also be used to obtain image representations for transfer learning. The basic pipeline for transfer learning is to first train a ConvNet on a large dataset (source task) and then use feed-forward units activation of the trained ConvNet as image representation for smaller datasets (target task). Our key contribution is to demonstrate superior performance of multiple ConvNet layer features over single ConvNet layer features. Combining multiple ConvNet layer features will result in more complex feature space with some features being repetitive. This requires some form of feature selection. We use AdaBoost with single stumps to implicitly select only distinct features that are useful towards classification from concatenated ConvNet features. Experimental results show that using multiple ConvNet layer activation features instead of single ConvNet layer features consistently will produce superior performance. Improvements becomes significant as we increase the distance between source task and the target task.

• 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.

• Elastomers and Composites
• /
• v.35 no.3
• /
• pp.173-179
• /
• 2000

Cure characteristics. tensile properties, and dynamic properties were investigated on the carbon black-filled natural rubber compounds, in which three typical vulcanization types conventional vulcanization(Conv), semi-efficient(Semi-EV), and efficient(EV) vulcanizations were used. The effects of vulcanization temperature on both the mechanical property and aging resistance of rubber compounds were also investigated. The Conv cure system showed a slightly slower rate of vulcanization than those of Semi-EV and EV ones. On the other hand, it showed a higher value in the maximum torque of cure curve. Higher tensile moduli were observed in Conv system than those in Semi-EV and EV ones, while lower elongation at break were obtained in Conv one. The tensile strength at break were found to be about the same for three cute systems. Hardness, modulus, and tensile strength decreased with increasing the vulcanization temperature, and the degree of changes in the properties was found to be smaller for EV and Semi-EV systems than that in Conv one. The EV system was found to be superior in thermal-aging resistance to Conv one.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.1
• /
• pp.55-62
• /
• 2019

Convolutional neural networks (CNNs) show high performance in the computer vision, but they require an enormous amount of operations, making them unsuitable for some resource- or energy-starving environments like the embedded environments. To overcome this problem, there have been much research on accelerators or pruning of CNNs. The previous pruning schemes have not considered the architecture of CNN accelerators, so the accelerators for the pruned CNNs have some inefficiency. This paper proposes a new pruning scheme, Conv-XP, which considers the architecture of CNN accelerators. In Conv-XP, the pruning is performed following the 'X' or '+' shape. The Conv-XP scheme induces a simple architecture of the CNN accelerators. The experimental results show that the Conv-XP scheme does not degrade the accuracy of CNNs, and that the accelerator area can be reduced by 12.8%.

• International Journal of Computer Science & Network Security
• /
• v.23 no.9
• /
• pp.47-54
• /
• 2023

For Bengali music emotion classification, deep learning models, particularly CNN and RNN are frequently used. But previous researches had the flaws of low accuracy and overfitting problem. In this research, attention-based Conv1D and BiGRU model is designed for music emotion classification and comparative experimentation shows that the proposed model is classifying emotions more accurate. We have proposed a Conv1D and Bi-GRU with the attention-based model for emotion classification of our Bengali music dataset. The model integrates attention-based. Wav preprocessing makes use of MFCCs. To reduce the dimensionality of the feature space, contextual features were extracted from two Conv1D layers. In order to solve the overfitting problems, dropouts are utilized. Two bidirectional GRUs networks are used to update previous and future emotion representation of the output from the Conv1D layers. Two BiGRU layers are conntected to an attention mechanism to give various MFCC feature vectors more attention. Moreover, the attention mechanism has increased the accuracy of the proposed classification model. The vector is finally classified into four emotion classes: Angry, Happy, Relax, Sad; using a dense, fully connected layer with softmax activation. The proposed Conv1D+BiGRU+Attention model is efficient at classifying emotions in the Bengali music dataset than baseline methods. For our Bengali music dataset, the performance of our proposed model is 95%.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1047-1056
• /
• 2022

Sea ice, frozen sea water, in the Artic is a primary indicator of global warming. Due to its importance to the climate system, shipping-route navigation, and fisheries, Arctic sea ice prediction has gained increased attention in various disciplines. Recent advances in artificial intelligence (AI), motivated by a desire to develop more autonomous and efficient future predictions, have led to the development of new sea ice prediction models as alternatives to conventional numerical and statistical prediction models. This study aims to evaluate the performance of the two-stream convolutional long-and short-term memory (TS-ConvLSTM) AI model, which is designed for learning both global and local characteristics of the Arctic sea ice changes, for the minimum September Arctic sea ice from 2001 to 2021, and to show the possibility for an operational prediction system. Although the TS-ConvLSTM model generally increased the prediction performance as training data increased, predictability for the marginal ice zone, 5-50% concentration, showed a negative trend due to increasing first-year sea ice and warming. Additionally, a comparison of sea ice extent predicted by the TS-ConvLSTM with the median Sea Ice Outlooks (SIOs) submitted to the Sea Ice Prediction Network has been carried out. Unlike the TS-ConvLSTM, the median SIOs did not show notable improvements as time passed (i.e., the amount of training data increased). Although the TS-ConvLSTM model has shown the potential for the operational sea ice prediction system, learning more spatio-temporal patterns in the difficult-to-predict natural environment for the robust prediction system should be considered in future work.

• Journal of Biomedical Engineering Research
• /
• v.43 no.5
• /
• pp.341-352
• /
• 2022

Medical image segmentation is the most important task in radiation therapy. Especially, when segmenting medical images, the liver is one of the most difficult organs to segment because it has various shapes and is close to other organs. Therefore, automatic segmentation of the liver in computed tomography (CT) images is a difficult task. Since tumors also have low contrast in surrounding tissues, and the shape, location, size, and number of tumors vary from patient to patient, accurate tumor segmentation takes a long time. In this study, we propose a method algorithm for automatically segmenting the liver and tumor for this purpose. As an advantage of setting the boundaries of the tumor, the liver and tumor were automatically segmented from the CT image using the 2D CoordConv DeepLab V3+ model using the CoordConv layer. For tumors, only cropped liver images were used to improve accuracy. Additionally, to increase the segmentation accuracy, augmentation, preprocess, loss function, and hyperparameter were used to find optimal values. We compared the CoordConv DeepLab v3+ model using the CoordConv layer and the DeepLab V3+ model without the CoordConv layer to determine whether they affected the segmentation accuracy. The data sets used included 131 hepatic tumor segmentation (LiTS) challenge data sets (100 train sets, 16 validation sets, and 15 test sets). Additional learned data were tested using 15 clinical data from Seoul St. Mary's Hospital. The evaluation was compared with the study results learned with a two-dimensional deep learning-based model. Dice values without the CoordConv layer achieved 0.965 ± 0.01 for liver segmentation and 0.925 ± 0.04 for tumor segmentation using the LiTS data set. Results from the clinical data set achieved 0.927 ± 0.02 for liver division and 0.903 ± 0.05 for tumor division. The dice values using the CoordConv layer achieved 0.989 ± 0.02 for liver segmentation and 0.937 ± 0.07 for tumor segmentation using the LiTS data set. Results from the clinical data set achieved 0.944 ± 0.02 for liver division and 0.916 ± 0.18 for tumor division. The use of CoordConv layers improves the segmentation accuracy. The highest of the most recently published values were 0.960 and 0.749 for liver and tumor division, respectively. However, better performance was achieved with 0.989 and 0.937 results for liver and tumor, which would have been used with the algorithm proposed in this study. The algorithm proposed in this study can play a useful role in treatment planning by improving contouring accuracy and reducing time when segmentation evaluation of liver and tumor is performed. And accurate identification of liver anatomy in medical imaging applications, such as surgical planning, as well as radiotherapy, which can leverage the findings of this study, can help clinical evaluation of the risks and benefits of liver intervention.

• Journal of KIISE
• /
• v.43 no.9
• /
• pp.1034-1042
• /
• 2016

Together with machine-learning frameworks, area-based imagery registration techniques can be easily applied to diverse types of image pairs without predefined features and feature descriptors. However, feature detectors are often used to quickly identify candidate image patch pairs, limiting the applicability of these registration techniques. In this paper, we propose a ConvNet (Convolutional Network) "Dart" that provides not only the matching metric between patches, but also information about their distance, which are helpful in reducing the search space of the corresponding patch pairs. In addition, we propose a ConvNet "Fad" to identify the patches that are difficult for Dart to improve the accuracy of registration. These two networks were successfully implemented using Deep Learning with the help of a number of training instances generated from a few registered image pairs, and were successfully applied to solve a simple image registration problem, suggesting that this line of research is promising.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2022.05a
• /
• pp.62-62
• /
• 2022

홍수 예보를 위한 강우-유출 분석에서 정확한 예측 강우량 정보는 매우 중요한 인자이다. 이에 따라 강우 예측을 위하여 다양한 연구들이 수행되고 있지만 시·공간적으로 비균일한 특성 또는 변동성을 가진 강우를 정확하게 예측하는 것은 여전히 난제이다. 본 연구에서는 딥러닝 기반 ConvLSTM (Convolutinal Long Short-Term Memory) 모형을 사용하여 위성 강수 자료의 단기 예측을 수행하고 그 정확성을 분석하고자 한다. 대상유역은 메콩강 유역이며, 유역 면적이 넓고 강우 관측소의 밀도가 낮아 시·공간적 강우량 추정에 한계가 있으므로 정확한 강우-유출 분석을 위하여 위성 강수 자료의 활용이 요구된다. 현재 TRMM, GSMaP, PERSIANN 등 많은 위성 강수 자료들이 제공되고 있으며, 우선적으로 ConvLSTM 모형의 강수 예측 활용가능성 평가를 위한 입력자료로 가장 보편적으로 활용되는 TRMM_3B42 자료를 선정하였다. 해당 자료의 특성으로 공간해상도는 0.25°, 시간해상도는 일자료이며, 2001년부터 2015년의 자료를 수집하였다. 모형의 평가를 위하여 2001년부터 2013년 자료는 학습, 2014년 자료는 검증, 2015년 자료는 예측에 사용하였다. 또한 민감도 분석을 통하여 ConvLSTM 모형의 최적 매개변수를 추정하고 이를 기반으로 선행시간(lead time) 1일, 2일, 3일의 위성 강수 예측을 수행하였다. 그 결과 선행시간이 길어질수록 그 오차는 증가하지만, 전반적으로 3가지 선행시간 모두 자료의 강수량뿐만 아니라 공간적 분포까지 우수하게 예측되었다. 따라서 2차원 시계열 자료의 특성을 기억하고 이를 예측에 반영할 수 있는 ConvLSTM 모형은 메콩강과 같은 미계측 대유역에서의 안정적인 예측 강수량 정보를 제공할 수 있으며 홍수 예보를 위한 강우-유출 분석에 활용이 가능할 것으로 판단된다.

• Korean Journal of Remote Sensing
• /
• v.36 no.5_3
• /
• pp.1077-1093
• /
• 2020

Sea Surface Temperature (SST) is an important environmental indicator that affects climate coupling systems around the world. In particular, coastal regions suffer from abnormal SST resulting in huge socio-economic damage. This study used Long Short Term Memory (LSTM) and Convolutional Long Short Term Memory (ConvLSTM) to predict SST up to 7 days in the south sea region in South Korea. The results showed that the ConvLSTM model outperformed the LSTM model, resulting in a root mean square error (RMSE) of 0.33℃ and a mean difference of -0.0098℃. Seasonal comparison also showed the superiority of ConvLSTM to LSTM for all seasons. However, in summer, the prediction accuracy for both models with all lead times dramatically decreased, resulting in RMSEs of 0.48℃ and 0.27℃ for LSTM and ConvLSTM, respectively. This study also examined the prediction of abnormally high SST based on three ocean heatwave categories (i.e., warning, caution, and attention) with the lead time from one to seven days for an ocean heatwave case in summer 2017. ConvLSTM was able to successfully predict ocean heatwave five days in advance.