• Journal of Korea Multimedia Society
• /
• v.25 no.10
• /
• pp.1375-1385
• /
• 2022

In this paper, we propose a semi-supervised domain adaptation solution to deal with practical face recognition (FR) scenarios where a single face image for each target identity (to be recognized) is only available in the training phase. Main goal of the proposed method is to reduce the discrepancy between the target and the source domain face images, which ultimately improves FR performances. The proposed method is based on the Domain Adatation network (DAN) using an MMD loss function to reduce the discrepancy between domains. In order to train more effectively, we develop a novel loss function learning strategy in which MMD loss and cross-entropy loss functions are adopted by using different weights according to the progress of each epoch during the learning. The proposed weight adoptation focuses on the training of the source domain in the initial learning phase to learn facial feature information such as eyes, nose, and mouth. After the initial learning is completed, the resulting feature information is used to training a deep network using the target domain images. To evaluate the effectiveness of the proposed method, FR performances were evaluated with pretrained model trained only with CASIA-webface (source images) and fine-tuned model trained only with FERET's gallery (target images) under the same FR scenarios. The experimental results showed that the proposed semi-supervised domain adaptation can be improved by 24.78% compared to the pre-trained model and 28.42% compared to the fine-tuned model. In addition, the proposed method outperformed other state-of-the-arts domain adaptation approaches by 9.41%.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.8
• /
• pp.3790-3803
• /
• 2018

Content-based image retrieval is an approach used to query images based on their semantics. Semantic based retrieval has its application in all fields including medicine, space, computing etc. Semantically generated binary hash codes can improve content-based image retrieval. These semantic labels / binary hash codes can be generated from unlabeled data using convolutional autoencoders. Proposed approach uses semi-supervised deep hashing with semantic learning and binary code generation by minimizing the objective function. Convolutional autoencoders are basis to extract semantic features due to its property of image generation from low level semantic representations. These representations of images are more effective than simple feature extraction and can preserve better semantic information. Proposed activation and loss functions helped to minimize classification error and produce better hash codes. Most widely used datasets have been used for verification of this approach that outperforms the existing methods.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.479-487
• /
• 2021

In this paper, we propose an Sound Event Detection (SED) model using self-training based on a noisy student model. The proposed SED model consists of two stages. In the first stage, a mean-teacher model based on an Residual Convolutional Recurrent Neural Network (RCRNN) is constructed to provide target labels regarding weakly labeled or unlabeled data. In the second stage, a self-training-based noisy student model is constructed by applying different noise types. That is, feature noises, such as time-frequency shift, mixup, SpecAugment, and dropout-based model noise are used here. In addition, a semi-supervised loss function is applied to train the noisy student model, which acts as label noise injection. The performance of the proposed SED model is evaluated on the validation set of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2020 Challenge Task 4. The experiments show that the single model and ensemble model of the proposed SED based on the noisy student model improve F1-score by 4.6 % and 3.4 % compared to the top-ranked model in DCASE 2020 challenge Task 4, respectively.

• Journal of Broadcast Engineering
• /
• v.27 no.2
• /
• pp.174-184
• /
• 2022

3D human pose estimation models can be classified into a multi-view model and a single-view model. In general, the multi-view model shows superior pose estimation performance compared to the single-view model. In the case of the single-view model, the improvement of the 3D pose estimation performance requires a large amount of training data. However, it is not easy to obtain annotations for training 3D pose estimation models. To address this problem, we propose a method to generate pseudo ground-truths of multi-view human pose data from a multi-view model and exploit the resultant pseudo ground-truths to train a single-view model. In addition, we propose a multi-view consistency loss function that considers the consistency of poses estimated from multi-view images, showing that the proposed loss helps the effective training of single-view models. Experiments using Human3.6M and MPI-INF-3DHP datasets show that the proposed method is effective for training single-view 3D human pose estimation models.

• Database Research
• /
• v.34 no.3
• /
• pp.137-147
• /
• 2018

A key to a well-performing human activity recognition (HAR) system through machine learning technique is the availability of a substantial amount of labeled data. Collecting sufficient labeled data is an expensive and time-consuming task. To build a HAR system in a new environment (i.e., the target domain) with very limited labeled data, it is unfavorable to naively exploit the data or trained classifier model from the existing environment (i.e., the source domain) as it is due to the domain difference. While traditional machine learning approaches are unable to address such distribution mismatch, transfer learning approach leverages the utilization of knowledge from existing well-established source domains that help to build an accurate classifier in the target domain. In this work, we propose a transfer learning approach to create an accurate HAR classifier with very limited data through the multitask neural network. The classifier loss function minimization for source and target domain are treated as two different tasks. The knowledge transfer is performed by simultaneously minimizing the loss function of both tasks using a single neural network model. Furthermore, we utilize the unlabeled data in an unsupervised manner to help the model training. The experiment result shows that the proposed work consistently outperforms existing approaches.