Earticle

Anomaly recognition in visual and audio data has gained increasing significance in computer vision, as it plays a crucial role in protecting human lives and property. In this work, we developed a semi-supervised multimodal framework for anomaly recognition that combines audio and visual data for better performance. The proposed framework employs a hybrid network consisting of a convolutional neural network, Bi-Directional Long Short-Term Memory, a multi-head attention module, and a fully connected layer for anomalous pattern recognition. We created a novel real-time visual-audio anomaly recognition dataset and evaluated our framework on it, achieving promising results.

의미론적 분할 형태로 합성곱 신경망을 구성하여 도로노면의 파손을 탐지하는 연구가 진행 되고 있다. 이러한 합성곱 신경망 형태의 모델을 생성하기 위해서는 입력 이미지와 이에 상응 한 레이블된 이미지 데이터셋으로 수집해야 하고, 이러한 과정에서는 굉장히 많은 시간과 비 용이 발생하게 된다. 본 논문에서는 이러한 작업을 완화하기 위하여 수도 레이블링을 활용한 준지도 학습 기반의 도로노면 파손 탐지 기술을 제안하고자 한다. 레이블된 데이터셋과 레이 블되지 않은 데이터셋을 적절하게 혼합하여 도로노면 파손을 탐지하는 모델을 업데이트하고, 이를 레이블된 데이터셋만을 활용한 기존 모델과 성능을 비교한다. 주관적인 성능결과, 민감도 부분에서는 조금 저하된 성능을 보였지만, 정밀도 부분에서는 대폭 성능 향상이 있었으며, 최 종적으로 F1-score 또한 높은 수치로 평가되었다

By using convolutional neural networks (CNNs) based on semantic segmentation, road surface damage detection has being studied. In order to generate the CNN model, it is essential to collect the input and the corresponding labeled images. Unfortunately, such collecting pairs of the dataset requires a great deal of time and costs. In this paper, we proposed a road surface damage detection technique based on semi-supervised learning using pseudo labels to mitigate such problem. The model is updated by properly mixing labeled and unlabeled datasets, and compares the performance against existing model using only labeled dataset. As a subjective result, it was confirmed that the recall was slightly degraded, but the precision was considerably improved. In addition, the F1-score was also evaluated as a high value.

Although living organisms differ in shape and size, all are fundamentally structured by genetic sequences. Interpreting these sequences helps explain how organisms function. With the advancement of AI, significant breakthroughs have been made in protein sequencing and understanding protein function. However, there is still room for improvement, as data-intensive models require a substantial amount of protein sequences, many of which are not publicly available or lack quality. In this paper, we present a semi-supervised learning scheme to address the shortage of high-quality training data necessary for training protein language models. We demonstrate that this approach enhances the model's capability to classify toxic fungi protein sequences.

Recently, due to the recent significant advances in machine learning and deep learning, it is being utilized in many fields. However, real-world data in the medical field significantly degrades the performance of machine learning algorithms due to problems that are heavily skewed to specific states or that the distribution of data is unbalanced. Therefore, this study solves the problem of not being learned by converting the dependent variable into a regression problem that predicts using a new dependent variable by pseudo labeling. Also, this study present ensemble methods to improve the performance of the model and prevent overfitting.

본 연구는 P2P(Peer-to-Peer) 대출의 부도위험 예측을 위하여 준지도학습(SSL) 기반의 모델을 개발하고자 한 다. 검증된 성능에도 불구하고 지도학습(SL) 방법은 완전 지불 또는 채무불이행과 같이 레이블이 결정된 다수의 데이터 가 필요한데 충분한 수의 레이블 데이터를 수집하려면 많은 자원과 시간이 필요하다. P2P 플랫폼이 급성장하면서 대출 건수도 매해 급증하였고, 레이블이 없는 데이터도 지속적으로 증가하고 있다. 본 연구는 P2P 대출 플랫폼인 LendingClub에서 수집한 데이터를 사용하였다. P2P 대출 중 레이블이 결정된 대출에서 추출한 정보뿐만 아니라 레이 블이 결정되지 않은 대출에서 추출한 정보도 사용하여 부도 위험을 예측하는 SSL 모델을 개발하여 연구를 수행한 결과, 적은 수의 레이블이 결정된 데이터를 사용함에도 불구하고 SSL 방법으로 구축된 모델이 많은 수의 레이블이 결정된 데이터를 사용하여 학습시킨 SL 방법으로 구축된 모델보다 부도 위험 예측성과가 향상되었다.

This study investigates the effect of the semi-supervised learning(SSL) method on predicting default risk of peer-to-peer(P2P) loans. Despite its proven performance, the supervised learning(SL) method requires labeled data, which may require a lot of effort and resources to collect. With the rapid growth of P2P platforms, the number of loans issued annually that have no clear final resolution is continuously increasing leading to abundance in unlabeled data. The research data of P2P loans used in this study were collected on the LendingClub platform. This is why an SSL model is needed to predict the default risk by using not only information from labeled loans(fully paid or defaulted) but also information from unlabeled loans. The results showed that in terms of default risk prediction and despite the use of a small number of labeled data, the SSL method achieved a much better default risk prediction performance than the SL method trained using a much larger set of labeled data.

의미론적 영상 분할을 위한 컨볼루션 신경망 기반 접근 방식은 픽셀 단위 레이블을 통한 지도 학습에 의존한다. 하 지만 접근 불가능한 도메인으로 일반화되지 않을 수 있는 문제점이 존재한다. 의미론적 영상 분할에서 모든 데이터 에 사람이 직접 레이블을 지정하는 작업은 노동 집약적이기 때문에 소량의 라벨링 데이터를 이용해 네트워크의 일반 화 성능을 향상시키는 것은 매우 중요하다. 본 논문에서는 가상 도메인과 현실 도메인 사이의 교차 도메인 혼합 샘 플링을 활용한 도메인 적응 학습 방법을 제안한다. 이를 위하여 혼합 샘플링 방법의 대표적인 기법을 분석하고 활용 하여 도메인 적응 문제에 적용하고 일반화 성능을 비교한다. 제안하는 학습 방법을 통해 학습된 네트워크는 도메인 적응 문제에서 정량적 수치와 정성적 결과 모두 기준 모델을 능가하는 성능을 보인다.

In semantic segmentation, a convolutional neural network-based approach relies on supervised learning through pixel-level labels. However, there is a problem that may not be generalized to other inaccessible domains. Therefore, it is very important to improve the generalization performance of the network using a small amount of labeling data because it is labor intensive to manually label all data. In this paper, we propose a domain adaptive learning method using cross-domain mixed sampling between the synthetic domain and the real-world domain. To this end, we analyze and utilize the representative techniques of the mixed sampling method, apply them to the domain adaptation problem, and compare the generalization performance. The network trained through the proposed learning method outperforms the baseline model in both quantitative and qualitative results in domain adaptation problems.

Semi-supervised learning (SSL) has shown great promise in utilizing both labeled and unlabeled data, especially using consistency regularization To enhance the model's learning efficiency, we proposed a dualbranch co-training approach. After dividing the training data into two subsets, each branch utilizes a teacherstudent model pair, where the teacher's weights are updated via an exponential moving average (EMA). The framework combines supervised loss and unsupervised loss to optimize the model. Upon label expansion (e.g pseudo labeling), an additional otherside view is introduced, promoting agreement between the branches' predictions on shared data. This loss mitigates errors arising from incorrect pseudo-labels and enhances the overall robustness of the training process. By dynamically adjusting pseudo-label inclusion based on confidence thresholds, our methodology reduces the impact of noisy data and prevents overfitting. As a result, we could demonstrate the effectiveness of the proposed method in leveraging unlabeled data while maintaining high performance.

소프트웨어 결함 예측 연구들의 대부분은 라벨 데이터를 훈련 데이터로 사용하는 감독형 모델에 관한 연구들이 다. 감독형 모델은 높은 예측 성능을 지니지만 대부분 개발 집단들은 충분한 라벨 데이터를 보유하고 있지 않다. 언라벨 데이터만 훈련에 사용하는 비감독형 모델은 모델 구축이 어렵고 성능이 떨어진다. 훈련 데이터로 라벨 데이터와 언라벨 데이터를 모두 사용하는 세미 감독형 모델은 이들의 문제점을 해결한다. Self-training은 세미 감독형 기법들 중 여러 가정과 제약조건들이 가장 적은 기법이다. 본 논문은 Self-training 알고리즘들을 이용해 여러 모델들을 구현하였으며, Accuracy와 AUC를 이용하여 그들을 평가한 결과 YATSI 모델이 가장 좋은 성능을 보였다.

Most studies of software fault prediction have been about supervised learning models that use only labeled training data. Although supervised learning usually shows high prediction performance, most development groups do not have sufficient labeled data. Unsupervised learning models that use only unlabeled data for training are difficult to build and show poor performance. Semi-supervised learning models that use both labeled data and unlabeled data can solve these problems. Self-training technique requires the fewest assumptions and constraints among semi-supervised techniques. In this paper, we implemented several models using self-training algorithms and evaluated them using Accuracy and AUC. As a result, YATSI showed the best performance.

매우 많은 소프트웨어 결함 예측에 관한 연구들이 수행되어왔지만 대부분은 라벨 데이터를 훈련 데이터로 사용하 는 감독형 모델들이었다. 언라벨 데이터만을 사용하는 비감독형 모델이나 언라벨 데이터와 매우 적은 라벨 데이터 정보 를 함께 사용하는 세미감독형 모델에 관한 연구는 극소수에 불과하다. 본 논문은 Self-training 기법에 트리 알고리즘들 을 사용하여 새로운 세미감독형 모델들을 제작하였다. 세미감독형 기법인 Self-training 모델에 트리 기법들을 사용하는 새로운 세미감독형 모델들을 제작하였다. 모델 평가 실험 결과 새롭게 제작한 트리 모델들이 기존 모델들보다 더 나은 성능을 보였으며, 특히 CollectiveWoods는 타 모델들에 비해 압도적으로 우월한 성능을 보였다. 또한 매우 적은 라벨 데이터 보유 상황에서도 매우 안정적인 성능을 보였다.

A number of studies have been conducted on predicting software faults, but most of them have been supervised models using labeled data as training data. Very few studies have been conducted on unsupervised models using only unlabeled data or semi-supervised models using enough unlabeled data and few labeled data. In this paper, we produced new semi-supervised models using tree algorithms in the self-training technique. As a result of the model performance evaluation experiment, the newly created tree models performed better than the existing models, and CollectiveWoods, in particular, outperformed other models. In addition, it showed very stable performance even in the case with very few labeled data.

최근 많은 이미지 데이터 셋들은 일반적인 특성을 추출하기 위한 다양한 데이터 클래스와 특징을 가지고 있다. 하지 만 이러한 다양한 데이터 클래스와 특징으로 인해 해당 데이터 셋으로 훈련된 물체 검출 딥러닝 모델은 데이터 특성이 다른 환경에서 좋은 성능을 내지 못하는 단점을 보인다. 이 논문에서는 하위 카테고리 기반 물체 검출 방법과 오픈셋 물체 검출 방법을 이용하여 이를 극복하고, 강인한 물체 검출 딥러닝 모델을 훈련하기 위해 능동 준지도 학습 (Active Semi-Supervised Learning) 을 이용한 다중 분기 트리 구조를 제안한다. 우리는 이 구조를 이용함으로써 데이터 특성이 다른 환경에서 적응 할 수 있는 모델을 가질 수 있고, 나아가 이 모델을 이용하여 이전의 모델보다 높은 성능을 확보 할 수 있다.

Recently there are many image datasets which has variety of data class and point to extract general features. But in order to this variety data class and point, deep learning model trained this dataset has not good performance in heterogeneous data feature local area. In this paper, we propose the structure which use sub-category and openset object detection methods to train more robust model, named multi-branch tree using ASSL. By using this structure, we can have more robust object detection deep learning model in heterogeneous data feature environment.

Automatic image annotation has been an active research topic in recent years due to its potential impact on both image understanding and semantic based image retrieval. However, the results of the state-of-the-art image annotation methods are still far from satisfaction due to the existence of semantic gap. Thus refining image annotation (RIA) has become one of the core research topics in computer vision and multimedia areas, whose purpose is to reserve the highly correlated annotations whereas remove the irrelevant or weakly relevant annotations by fully exploring the correlations of annotation keywords. RIA, to some extent, can effectively mitigate the semantic gap between low-level visual features and high-level semantic concepts. So in this paper, we focus on the latest development in image retrieval and provide a comprehensive survey on refining image annotation techniques. In particular, we analyze the key aspects of various RIA methods, including their original intentions and annotation models. Finally, we draw some important conclusions and highlight the potential research directions for the future.

Tracking with a discriminative classifier becomes popular recently. The online updating makes it easy to adapt to target appearance variations. However, this also brings drifting problem. It’s necessary to find a tracking method with strong adaptivity and anti-drifting ability. In this paper, an online semi-supervised boosting method is proposed at first, and based on it, we propose a novel tracking framework that treats samples differently when updating the classifier under different conditions. This tracking framework can significantly alleviate the drifting problem and keep adaptive enough to appearance variations. Experimental results on challenging videos show that our method can track accurately and robustly， and outperform many other state-of-the-art trackers.

In recent years, semi-supervised learning has been a hot research topic in machine learn-ing area. Different from traditional supervised learning which learns only from labeled data; semi-supervised learning makes use of both labeled and unlabeled data for learning purpose. Co-training is a popular semi-supervised learning algorithm which assumes that each exam-ple is represented by two or more redundantly sufficient sets of features (views) and addi-tionally these views are independent given the class. To improve the performance and ap-plicability of co-training, ensemble learning, such as bagging and random subspace has been used along with co-training. In this work, we propose to use the rough set based ensem-ble learning method with co-training algorithm (RSCO). Inherited the inherent characteris-tics of rough set, ensemble learning is expected to meet both the diversity and accuracy re-quirement. Finally experimental results on the UCI data sets demonstrate the promising per-formance of RSCO.

Recently, the license plate OCR system has been commercialized in a variety of fields and preferred utilizing low-cost embedded systems using only cameras. This system has a high recognition rate of about 98% or more for the environments such as parking lots where non-vehicle is restricted; however, the environments where non-vehicle objects are not restricted, the recognition rate is about 50% to 70%. This low performance is due to the changes in the environment by non-vehicle objects in real-time situations that occur anomaly data which is similar to the license plates. In this paper, we implement the appropriate anomaly detection based on semisupervised learning for the license plate OCR system in the real-time environment where the appearance of non-vehicle objects is not restricted. In the experiment, we compare systems which anomaly detection is not implemented in the preceding research with the proposed system in this paper. As a result, the systems which anomaly detection is not implemented had a recognition rate of 77%; however, the systems with the semisupervised learning based on anomaly detection had 88% of recognition rate. Using the techniques of anomaly detection based on the semi-supervised learning was effective in detecting anomaly data and it was helpful to improve the recognition rate of real-time situations.

In this paper, we present a new method for automatic image annotation by applying semi- supervised learning based on the Bayesian framework. On the one hand, we employ the semi- supervised learning, i.e., transductive support vector machine (TSVM) to enhance the quality of training image data, which is a promising way to find out the underlying relevant data from the unlabeled ones. On the other hand, a simple yet very efficient Bayesian model is built to implement image annotation by the maximum a posteriori (MAP) criterion. The novelty of our method mainly lies in two aspects: exploiting TSVM to improve the quality of training image dataset and utilizing the Bayesian model to predict the candidate annotations for the unseen images. Experimental results on the standard Corel dataset demonstrate that the proposed method is superior or highly competitive to several state-of-the-art approaches.

The real challenge in pattern recognition tasks and machine learning processes is to train a discriminator using labeled data and use it to distinguish between future data points as accurate as possible. However, most of the problems in the real world have numerous data. Therefore assigning labels to every data points in these problems are a cumbersome or even impossible matter. Semi-supervised learning is one approach to overcome these types of problems. It uses only a small set of labeled with the company of huge remain and unlabeled data to train the discriminator. In semi-supervised learning, it is very essential that which data is labeled and depend on position of data it effectiveness changes. In this paper, we proposed an evolutionary approach called Artificial Immune System (AIS) to determine which data is better to be labeled to get the high quality data. The experimental results represent the effectiveness of this algorithm in finding these data points.

Short text is a popular text form, which is widely used in real-time network news, short commentary, micro-blog and many other fields. With the development of the application such as QQ, mobile phone text messages and movie websites, the size of data is also becoming larger and larger. Most data is useless for us while other data is significant for us. Therefore, it is necessary for us to extract the useful short text from the big data. However, there are many problems with the short text classification, such as fewer features, irregularity and so on. To solve these problems, we should pretreat the short text set first, and then choose the significant features. This paper use semi-supervised learning method and SVM classifier to improve the traditional methods and it can classify a large number of short texts to mining the useful massage from the short text. The experimental results in this paper also show a good promotion.

사람의 얼굴 표정을 실제 환경에서 인식하는 데에는 여러 가지 난이한 점이 존재한다. 그래서 학습에 사용된 데이터베이스와 실험 데이터가 여러 가지 조건이 비슷할 때에만 그 성능이 높게 나온다. 이러한 문제점을 해결하려면 수많은 얼굴 표정 데이터가 필요하다. 본 논문에서는 능동준지도 학습을 통해 다양한 조건의 얼굴 표정 데이터를 쉽게 모으고 보다 빠르게 성능을 확보할 수 있는 방법을 제안한다. 제안하는 알고리즘은 딥러닝 네트워크와 능동 학습 (Active Learning)을 통해 초기 모델을 학습하고, 이후로는 준지도 학습(Semi-Supervised Learning)을 통해 라벨이 없는 추가 데이터를 확보하며, 성능이 확보될 때까지 이러한 과정을 반복한다. 위와 같은 능동준지도 학습(Active Semi-Supervised Learning)을 통해서 보다 적은 노동력으로 다양한 환경에 적합한 데이터를 확보하여 성능을 확보할 수 있다.

It is difficult to recognize Human’s facial expression in the real-world. For these reason, when database and test data have similar condition, we can accomplish high accuracy. Solving these problem, we need to many facial expression data. In this paper, we propose the algorithm for gathering many facial expression data within various environment and gaining high accuracy quickly. This algorithm is training initial model with the ASSL (Active Semi-Supervised Learning) using deep learning network, thereafter gathering unlabeled facial expression data and repeating this process. Through using the ASSL, we gain proper data and high accuracy with less labor force.

본 연구는 GAN(Generative Adversarial Network)을 기반으로 한 Semi Supervised Learning을 활용하여 이미지 생성과 ResNet50을 이용한 이미지 분류를 결합하는 방법에 대해 다루고 있다. 이를 통해 새로운 접근법을 제시하여 이미지 생성과 분류를 통합함으로써 더 정확하고 다양한 결과를 얻을 수 있도록 하였다. 생성자와 판별자를 학습시켜 생성된 이미지와 실제 이미지를 구별하고, ResNet50을 활용하여 이미지 분류를 수행한다. 실험 결과에서는 생성된 이미지의 품질이 epoch에 따라 변화함을 확인할 수 있었으며, 이를 통해 산업재해 예측 정확성을 향상하고자 한다. 또한, GAN과 ResNet50의 결합을 통해 이미지 생성의 품질을 향상시키고 이미지 분류의 정확도를 높이는 효율적인 방법을 제시하고자 한다.

This study deals with a method of combining image generation using Semi Supervised Learning based on GAN (Generative Adversarial Network) and image classification using ResNet50. Through this, a new approach was proposed to obtain more accurate and diverse results by integrating image generation and classification. The generator and discriminator are trained to distinguish generated images from actual images, and image classification is performed using ResNet50. In the experimental results, it was confirmed that the quality of the generated images changes depending on the epoch, and through this, we aim to improve the accuracy of industrial accident prediction. In addition, we would like to present an efficient method to improve the quality of image generation and increase the accuracy of image classification through the combination of GAN and ResNet50.