• International Journal of Computer Science & Network Security
• /
• v.21 no.7
• /
• pp.182-190
• /
• 2021

Deep learning is an advanced technology for large-scale data analysis, with numerous promising cases like image processing, object detection and significantly more. It becomes customarily to use transfer learning and fine-tune a pre-trained CNN model for most image recognition tasks. Having people taking photos and tag themselves provides a valuable resource of in-data. However, these tags and labels might be noisy as people who annotate these images might not be experts. This paper aims to explore the impact of noisy labels on fine-tuning pre-trained CNN models. Such effect is measured on a food recognition task using Food101 as a benchmark. Four pre-trained CNN models are included in this study: InceptionV3, VGG19, MobileNetV2 and DenseNet121. Symmetric label noise will be added with different ratios. In all cases, models based on DenseNet121 outperformed the other models. When noisy labels were introduced to the data, the performance of all models degraded almost linearly with the amount of added noise.

• Journal of Korea Multimedia Society
• /
• v.23 no.11
• /
• pp.1349-1360
• /
• 2020

In recent years, Convolutional Neural Networks (CNNs) have been successfully implemented in different tasks of computer vision. Since CNN models are the representatives of supervised learning algorithms, they demand large amount of data in order to train the classifiers. Thus, obtaining data with correct labels is imperative to attain the state-of-the-art performance of the CNN models. However, labelling datasets is quite tedious and expensive process, therefore real-life datasets often exhibit incorrect labels. Although the issue of poorly labelled datasets has been studied before, we have noticed that the methods are very complex and hard to reproduce. Therefore, in this research work, we propose Deep CleanNet - a considerably simple system that achieves competitive results when compared to the existing methods. We use K-means clustering algorithm for selecting data with correct labels and train the new dataset using a deep CNN model. The technique achieves competitive results in both training and validation stages. We conducted experiments using MNIST database of handwritten digits with 50% corrupted labels and achieved up to 10 and 20% increase in training and validation sets accuracy scores, respectively.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.5
• /
• pp.2523-2538
• /
• 2017

In most of the existing literature, the definition of the class label has the following characteristics. First, the class label of the samples from the same object has an absolutely fixed value. Second, the difference between class labels of the samples from different objects should be maximized. However, the appearance of a face varies greatly due to the variations of the illumination, pose, and expression. Therefore, the previous definition of class label is not quite reasonable. Inspired by discriminative least squares regression algorithm (DLSR), a noisy label based discriminative least squares regression algorithm (NLDLSR) is presented in this paper. In our algorithm, the maximization difference between the class labels of the samples from different objects should be satisfied. Meanwhile, the class label of the different samples from the same object is allowed to have small difference, which is consistent with the fact that the different samples from the same object have some differences. In addition, the proposed NLDLSR is expanded to the kernel space, and we further propose a novel kernel noisy label based discriminative least squares regression algorithm (KNLDLSR). A large number of experiments show that our proposed algorithms can achieve very good performance.

• International Journal of Fuzzy Logic and Intelligent Systems
• /
• v.14 no.1
• /
• pp.8-16
• /
• 2014

We consider the visual barcode recognition problem in a noisy video data setup. Unlike most existing single-frame recognizers that require considerable user effort to acquire clean, motionless and blur-free barcode signals, we eliminate such extra human efforts by proposing a robust video-based barcode recognition algorithm. We deal with a sequence of noisy blurred barcode image frames by posing it as an online filtering problem. In the proposed dynamic recognition model, at each frame we infer the blur level of the frame as well as the digit class label. In contrast to a frame-by-frame based approach with heuristic majority voting scheme, the class labels and frame-wise noise levels are propagated along the frame sequences in our model, and hence we exploit all cues from noisy frames that are potentially useful for predicting the barcode label in a probabilistically reasonable sense. We also suggest a visual barcode tracking approach that efficiently localizes barcode areas in video frames. The effectiveness of the proposed approaches is demonstrated empirically on both synthetic and real data setup.

• Journal of Communications and Networks
• /
• v.5 no.2
• /
• pp.157-166
• /
• 2003

This paper deals with the problem of channel identification for Single Input Multiple Output (SIMO) slow fading channels using clustering algorithms. Due to the intrinsic memory of the discrete-time model of the channel, over short observation periods, the received data vectors of the SIMO model are spread in clusters because of the AWGN noise. Each cluster is practically centered around the ideal channel output labels without noise and the noisy received vectors are distributed according to a multivariate Gaussian distribution. Starting from the Markov SIMO channel model, simultaneous maximum ikelihood estimation of the input vector and the channel coefficients reduce to one of obtaining the values of this pair that minimizes the sum of the Euclidean norms between the received and the estimated output vectors. Viterbi algorithm can be used for this purpose provided the trellis diagram of the Markov model can be labeled with the noiseless channel outputs. The problem of identification of the ideal channel outputs, which is the focus of this paper, is then equivalent to designing a Vector Quantizer (VQ) from a training set corresponding to the observed noisy channel outputs. The Linde-Buzo-Gray (LBG)-type clustering algorithms [1] could be used to obtain the noiseless channel output labels from the noisy received vectors. One problem with the use of such algorithms for blind time-varying channel identification is the codebook initialization. This paper looks at two critical issues with regards to the use of VQ for channel identification. The first has to deal with the applicability of this technique in general; we present theoretical results for the conditions under which the technique may be applicable. The second aims at overcoming the codebook initialization problem by proposing a novel approach which attempts to make the first phase of the channel estimation faster than the classical codebook initialization methods. Sample simulation results are provided confirming the effectiveness of the proposed initialization technique.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.05a
• /
• pp.607-610
• /
• 2021

When learning a model in supervised learning, input data and the label of the data are required. However, labeling is high cost task and if automated, there is no guarantee that the label will always be correct. In the case of supervised learning in such a noisy labels environment, the accuracy of the model increases at the initial stage of learning, but decrease significantly after a certain period of time. There are various methods to solve the noisy label problem. But in most cases, the probability predicted by the model is used as the pseudo label. So, we proposed a method to predict the true label more quickly by refining the probabilities predicted by the model. Result of experiments on the same environment and dataset, it was confirmed that the performance improved and converged faster. Through this, it can be applied to methods that use the probability distribution predicted by the model among existing studies. And it is possible to reduce the time required for learning because it can converge faster in the same environment.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.12
• /
• pp.4776-4794
• /
• 2020

We propose a rapid adaptive learning framework for streaming object detection, called EER-ASSL. The method combines the expected error reduction (EER) dependent rollback learning and the active semi-supervised learning (ASSL) for a rapid adaptive CNN detector. Most CNN object detectors are built on the assumption of static data distribution. However, images are often noisy and biased, and the data distribution is imbalanced in a real world environment. The proposed method consists of collaborative sampling and EER-ASSL. The EER-ASSL utilizes the active learning (AL) and rollback based semi-supervised learning (SSL). The AL allows us to select more informative and representative samples measuring uncertainty and diversity. The SSL divides the selected streaming image samples into the bins and each bin repeatedly transfers the discriminative knowledge of the EER and CNN models to the next bin until convergence and incorporation with the EER rollback learning algorithm is achieved. The EER models provide a rapid short-term myopic adaptation and the CNN models an incremental long-term performance improvement. EER-ASSL can overcome noisy and biased labels in varying data distribution. Extensive experiments shows that EER-ASSL obtained 70.9 mAP compared to state-of-the-art technology such as Faster RCNN, SSD300, and YOLOv2.

• 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 Information Processing Systems
• /
• v.11 no.4
• /
• pp.528-537
• /
• 2015

Text in images is one of the most important cues for understanding a scene. In this paper, we propose a novel approach based on interest points to localize text in natural scene images. The main ideas of this approach are as follows: first we used interest point detection techniques, which extract the corner points of characters and center points of edge connected components, to select candidate regions. Second, these candidate regions were verified by using tensor voting, which is capable of extracting perceptual structures from noisy data. Finally, area, orientation, and aspect ratio were used to filter out non-text regions. The proposed method was tested on the ICDAR 2003 dataset and images of wine labels. The experiment results show the validity of this approach.

• Journal of KIISE
• /
• v.44 no.9
• /
• pp.887-892
• /
• 2017

Brain-computer interface (BCI) is a technology that controls computer and transmits intention by measuring and analyzing electroencephalogram (EEG) signals generated in multi-channel during mental work. At this time, optimal EEG channel selection is necessary not only for convenience and speed of BCI but also for improvement in accuracy. The optimal channel is obtained by removing duplicate(redundant) channels or noisy channels. This paper propose a dual filter-based channel selection method to select the optimal EEG channel. The proposed method first removes duplicate channels using Spearman's rank correlation to eliminate redundancy between channels. Then, using F score, the relevance between channels and class labels is obtained, and only the top m channels are then selected. The proposed method can provide good classification accuracy by using features obtained from channels that are associated with class labels and have no duplicates. The proposed channel selection method greatly reduces the number of channels required while improving the average classification accuracy.