• Journal of Broadcast Engineering
• /
• v.27 no.2
• /
• pp.185-197
• /
• 2022

Recently, deep learning networks with high performance for object recognition are emerging. In the case of object recognition using deep learning, it is important to build a training data set to improve performance. To build a data set, we need to collect and label the images. This process requires a lot of time and manpower. For this reason, open data sets are used. However, there are objects that do not have large open data sets. One of them is data required for license plate detection and recognition. Therefore, in this paper, we propose an artificial license plate generator system that can create large data sets by minimizing images. In addition, the detection rate according to the artificial license plate arrangement structure was analyzed. As a result of the analysis, the best layout structure was FVC_III and B, and the most suitable network was D2Det. Although the artificial data set performance was 2-3% lower than that of the actual data set, the time to build the artificial data was about 11 times faster than the time to build the actual data set, proving that it is a time-efficient data set building system.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.41 no.8
• /
• pp.958-964
• /
• 2016

In this paper, we propose an i-vector based language recognition system to identify the spoken language of the speaker, which uses multiple discriminative deep neural network (DNN) models analogous to the multi-class support vector machine (SVM) classification system. The proposed model was trained and tested using the i-vectors included in the NIST 2015 i-vector Machine Learning Challenge database, and shown to outperform the conventional language recognition methods such as cosine distance, SVM and softmax NN classifier in open-set experiments.

• Proceedings of the IEEK Conference
• /
• 2003.07e
• /
• pp.1871-1874
• /
• 2003

Knowledge-based numeric open caption recognition is proposed that can recognize numeric captions generated by character generator (CG) and automatically superimpose a modified caption using the recognized text only when a valid numeric caption appears in the aimed specific region of a live sportscast scene produced by other broadcasting stations. in the proposed method, mesh features are extracted from an enhanced binary image as feature vectors, then a valuable information is recovered from a numeric image by perceiving the character using a multiplayer perceptron (MLP) network. The result is verified using knowledge-based hie set designed for a more stable and reliable output and then the modified information is displayed on a screen by CG. MLB Eye Caption based on the proposed algorithm has already been used for regular Major League Base-ball (MLB) programs broadcast five over a Korean nationwide TV network and has produced a favorable response from Korean viewer.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.83-88
• /
• 2023

This study investigates the fine-tuning of large-scale Automatic Speech Recognition (ASR) models, specifically OpenAI's Whisper model, for domain-specific applications using the KsponSpeech dataset. The primary research questions address the effectiveness of targeted lexical item emphasis during fine-tuning, its impact on domain-specific performance, and whether the fine-tuned model can maintain generalization capabilities across different languages and environments. Experiments were conducted using two fine-tuning datasets: Set A, a small subset emphasizing specific lexical items, and Set B, consisting of the entire KsponSpeech dataset. Results showed that fine-tuning with targeted lexical items increased recognition accuracy and improved domain-specific performance, with generalization capabilities maintained when fine-tuned with a smaller dataset. For noisier environments, a trade-off between specificity and generalization capabilities was observed. This study highlights the potential of fine-tuning using minimal domain-specific data to achieve satisfactory results, emphasizing the importance of balancing specialization and generalization for ASR models. Future research could explore different fine-tuning strategies and novel technologies such as prompting to further enhance large-scale ASR models' domain-specific performance.

• Journal of IKEEE
• /
• v.22 no.4
• /
• pp.1131-1139
• /
• 2018

This paper proposes a method of recognition abandoned objects using convolutional neural networks. The method first detects an area for an abandoned object in image and, if there is a detected area, applies convolutional neural networks to that area to recognize which object is represented. Experiments were conducted through an application system that detects illegal trash dumping. The experiments result showed the area of abandoned object was detected efficiently. The detected areas enter the input of convolutional neural networks and are classified into whether it is a trash or not. To do this, I trained convolutional neural networks with my own trash dataset and open database. As a training result, I achieved high accuracy for the test set not included in the training set.

• ETRI Journal
• /
• v.28 no.6
• /
• pp.811-814
• /
• 2006

In this letter, we introduce confusion-based confidence measures for detecting an impostor in speaker recognition, which does not require an alternative hypothesis. Most traditional speaker verification methods are based on a hypothesis test, and their performance depends on the robustness of an alternative hypothesis. Compared with the conventional Gaussian mixture model-universal background model (GMM-UBM) scheme, our confusion-based measures show better performance in noise-corrupted speech. The additional computational requirements for our methods are negligible when used to detect or reject impostors.

• Proceedings of the KSPS conference
• /
• 2007.05a
• /
• pp.37-38
• /
• 2007

This paper reviews the issues in implementing sound recognizers in real environments. First is the signal corruption caused by background noises and reverberation. Second is the open-set problem which is the problem of rejecting out-of-vocabulary words and noises. These two issues must be solved for noise robust recognizers.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.32B no.6
• /
• pp.909-921
• /
• 1995

The objective of this paper is to develop an algorithm for a real-time automatic fingerprint recognition system. The algorithm employs the Fast Fourier Transform (FFT) in determining the directions of ridges in fingerprint images, and utilizes statistical information in recognizing the fingerprints. The information used in fingerprint recognition is based on the dircetions along ridge curves and characteristic points such as core points and delta points. In order to find ridge directions, the algorithm applies the FFT to a small block of the size 8x8 pixels, and decides the directions by interpreting the resulted Fourier spectrum. By using the FFT, the algorithm does not require conventional preprocessing procedures such as smoothing, binarization, thinning, and restorationl. Finally, in matching two fingerprint images, the algorithm searches and compares two kinds of feature blocks, one as the blocks where the dircetions cannot be defined from the Fourier spectrum, and the other as the blocks where the changes of directions become abrupt. The proposed algorithm has been implemented on a SunSparc-2 workstation under the Open Window environment. In the experiment, the proposed algorithm has been applied to a set of fingerprint images obtained by a prism system. The result has shown that while the rate of Type II error - Incorrect recognition of two different fingerprints as the identical fingerprints - is held at 0.0%, the rate of Type I error - Incorrect recognition of two identical fingerprints as the different ones - is 2.2%.

• Smart Media Journal
• /
• v.13 no.7
• /
• pp.36-44
• /
• 2024

Identifying plant species and diseases is crucial for maintaining biodiversity and achieving optimal crop yields, making it a topic of significant practical importance. Recent studies have extended plant disease recognition from traditional closed-set scenarios to open-set environments, where the goal is to reject samples that do not belong to known categories. However, in open-world tasks, it is essential not only to define unknown samples as "unknown" but also to classify them further. This task assumes that images and labels of known categories are available and that samples of unknown categories can be accessed. The model classifies unknown samples by learning the prior knowledge of known categories. To the best of our knowledge, there is no existing research on this topic in plant-related recognition tasks. To address this gap, this paper utilizes knowledge distillation to model the category space relationships between known and unknown categories. Specifically, we identify similarities between different species or diseases. By leveraging a fine-tuned model on known categories, we generate pseudo-labels for unknown categories. Additionally, we enhance the baseline method's performance by using a larger pre-trained model, dino-v2. We evaluate the effectiveness of our method on the large plant specimen dataset Herbarium 19 and the disease dataset Plant Village. Notably, our method outperforms the baseline by 1% to 20% in terms of accuracy for novel category classification. We believe this study will contribute to the community.

• International Journal of Computer Science & Network Security
• /
• v.21 no.3
• /
• pp.37-40
• /
• 2021

This paper presents a design of a speech to sign language converter for deaf and hard of hearing people. The device is low-cost, low-power consumption, and it can be able to work entirely offline. The speech recognition is implemented using an open-source API, Pocketsphinx library. In this work, we proposed a context-oriented language model, which measures the similarity between the recognized speech and the predefined speech to decide the output. The output speech is selected from the recommended speech stored in the database, which is the best match to the recognized speech. The proposed context-oriented language model can improve the speech recognition rate by 21% for working entirely offline. A decision module based on determining the similarity between the two texts using Levenshtein distance decides the output sign language. The output sign language corresponding to the recognized speech is generated as a set of sequential images. The speech to sign language converter is deployed on a Raspberry Pi Zero board for low-cost deaf assistive devices.