• MALSORI
• /
• no.60
• /
• pp.165-180
• /
• 2006

In this paper, we present several confidence measures (CM) for speech recognition systems to evaluate the reliability of recognition results. We propose heuristic CMs such as mean log-likelihood score, N-best word log-likelihood ratio, likelihood sequence fluctuation and likelihood ratio testing(LRT)-based CMs using several types of anti-models. Furthermore, we propose new algorithms to add weighting terms on phone-level log-likelihood ratio to merge word-level log-likelihood ratios. These weighting terms are computed from the distance between acoustic models and knowledge-based phoneme classifications. LRT-based CMs show better performance than heuristic CMs excessively, and LRT-based CMs using phonetic information show that the relative reduction in equal error rate ranges between $8{\sim}13%$ compared to the baseline LRT-based CMs. We use the support vector machine to fuse several CMs and improve the performance of utterance verification. From our experiments, we know that selection of CMs with low correlation is more effective than CMs with high correlation.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.10
• /
• pp.4808-4824
• /
• 2016

Signal individual reconnaissance and identification is an extremely important research topic in non-cooperative domains such as electronic countermeasures and intelligence reconnaissance. Facing the characteristics of the complexity and changeability of current communication environment, how to realize radiation source signal individual identification under the low SNR conditions is an emphasis of research. A novel emitter individual identification method combined bi-spectrum analysis with wavelet feature is presented in this paper. It makes a feature fusion of bi-spectrum slice characteristics and energy variance characteristics of the secondary wavelet transform coefficient to identify MFSK signals under the low SNR (signal-to-noise ratios) environment. Theoretical analyses and computer simulation results show that the proposed algorithm has good recognition performance with the ability to suppress noise and interference, and reaches the recognition rate of more than 90% when the SNR is -6dB.

• Annual Conference of KIPS
• /
• 2015.10a
• /
• pp.1711-1714
• /
• 2015

스마트 디바이스가 사회와 소통할 수 있는 도구가 되었음에도 불구하고 아직까지 노인들이 사용하기에는 어려움이 있다. 여기에 음성인식 기술을 이용한 음성인터페이스를 활용함으로써 노인들의 스마트 디바이스에 대한 사용성을 높일 수 있다. 하지만 일반적인 음성인식 시스템은 청장년의 발성 스타일에 맞춰져 있기 때문에, 노화된 노인의 발성이 그대로 입력될 경우 음성인식률이 하락한다. 본 연구에서는 노인의 음절 별 발화속도가 일반적인 음성인식 시스템의 성능을 보증할 수 있는 범위를 벗어나는 경우가 많다는 분석 결과를 토대로 노인의 음절 별 발화속도를 조정한 결과 노인남녀 평균 음성인식률이 15.3% 상승하였다. 이처럼 노인의 음성인식 오류 원인들 중 하나인 발화속도의 재조정으로 음성 인식률을 높일 수 있는 토대를 마련하였다. 이는 노인들이 스마트 디바이스를 이용하여 쉽고 정확한 작업을 수행할 수 있게 됨으로써, 노인들의 사회 참여와 정보 획득이 용이해 지고 더 나아가 세대 간의 소통에도 이바지할 것으로 기대한다.

• Journal of the Society of Naval Architects of Korea
• /
• v.58 no.2
• /
• pp.105-111
• /
• 2021

The main source of underwater radiated noise of ships is cavitation generated by propeller blades. After the Cavitation Inception Speed (CIS), noise level at all frequencies increases severely. In determining the CIS, it is based on the results observed with the naked eye during the model test, however accuracy and consistency of CIS values are becoming practical issues. This study was carried out with the aim of developing a technology that can automatically recognize cavitation images using deep learning technique based on a Convolutional Neural Network (CNN). Model tests on a three-dimensional hydrofoil were conducted at a cavitation tunnel, and tip vortex cavitation was strictly observed using a high-speed camera to obtain analysis data. The results show that this technique can be used to quantitatively evaluate not only the CIS, but also the amount and rate of cavitation from recorded images.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.1
• /
• pp.131-146
• /
• 2021

Multimodal biometric-based recognition has been an active topic because of its higher convenience in recent years. Due to high user convenience of finger, finger-based personal identification has been widely used in practice. Hence, taking Finger-Print (FP), Finger-Vein (FV) and Finger-Knuckle-Print (FKP) as the ingredients of characteristic, their feature representation were helpful for improving the universality and reliability in identification. To usefully fuse the multimodal finger-features together, a new robust representation algorithm was proposed based on hierarchical model. Firstly, to obtain more robust features, the feature maps were obtained by Gabor magnitude feature coding and then described by Local Binary Pattern (LBP). Secondly, the LGBP-based feature maps were processed hierarchically in bottom-up mode by variable rectangle and circle granules, respectively. Finally, the intension of each granule was represented by Local-invariant Gray Features (LGFs) and called Hierarchical Local-Gabor-based Gray Invariant Features (HLGGIFs). Experiment results revealed that the proposed algorithm is capable of improving rotation variation of finger-pose, and achieving lower Equal Error Rate (EER) in our homemade database.

• International Journal of Computer Science & Network Security
• /
• v.21 no.5
• /
• pp.256-266
• /
• 2021

The Convolutional Neural Network (CNN) has recently made potential improvements in face verification applications. In fact, different models based on the CNN have attained commendable progress in the classification rate using a massive amount of data in an uncontrolled environment. However, the enormous computation costs and the considerable use of storage causes a noticeable problem during training. To address these challenges, we focus on relevant data trained within the CNN model by integrating a lifting method for a better tradeoff between the data size and the computational efficiency. Our approach is characterized by the advantage that it does not need any additional space to store the features. Indeed, it makes the model much faster during the training and classification steps. The experimental results on Labeled Faces in the Wild and YouTube Faces datasets confirm that the proposed CNN framework improves performance in terms of precision. Obviously, our model deliberately designs to achieve significant speedup and reduce computational complexity in deep CNNs without any accuracy loss. Compared to the existing architectures, the proposed model achieves competitive results in face recognition tasks

• Journal of Korea Multimedia Society
• /
• v.24 no.3
• /
• pp.416-422
• /
• 2021

Due to the large differences in human activity within classes, the large similarity between classes, and the problems of visual angle and occlusion, it is difficult to extract features manually, and the detection rate of human behavior is low. In order to better solve these problems, an improved Faster R-CNN-based detection algorithm is proposed in this paper. It achieves multi-object recognition and localization through a second-order detection network, and replaces the original feature extraction module with Dense-Net, which can fuse multi-level feature information, increase network depth and avoid disappearance of network gradients. Meanwhile, the proposal merging strategy is improved with Soft-NMS, where an attenuation function is designed to replace the conventional NMS algorithm, thereby avoiding missed detection of adjacent or overlapping objects, and enhancing the network detection accuracy under multiple objects. During the experiment, the improved Faster R-CNN method in this article has 84.7% target detection result, which is improved compared to other methods, which proves that the target recognition method has significant advantages and potential.

• International Journal of Advanced Culture Technology
• /
• v.10 no.1
• /
• pp.230-235
• /
• 2022

This study is to select an optimal object detection algorithm for designing a self-checkout counter to improve the inconvenience of payment systems for products without existing barcodes. To this end, a performance comparison analysis of YOLO v2, Tiny YOLO v2, and the latest YOLO v5 among deep learning-based object detection algorithms was performed to derive results. In this paper, performance comparison was conducted by forming learning data as an example of 'donut' in a bakery store, and the performance result of YOLO v5 was the highest at 96.9% of mAP. Therefore, YOLO v5 was selected as the artificial intelligence object detection algorithm to be applied in this paper. As a result of performance analysis, when the optimal threshold was set for each donut, the precision and reproduction rate of all donuts exceeded 0.85, and the majority of donuts showed excellent recognition performance of 0.90 or more. We expect that the results of this paper will be helpful as the fundamental data for the development of an automatic payment system using AI self-service technology that is highly usable in the non-face-to-face era.

• Journal of Biomedical Engineering Research
• /
• v.43 no.6
• /
• pp.441-447
• /
• 2022

In this paper, we propose a pill identification model using engraved text feature and image feature such as shape and color, and compare it with an identification model that does not use engraved text feature to verify the possibility of improving identification performance by improving recognition rate of the engraved text. The data consisted of 100 classes and used 10 images per class. The engraved text feature was acquired through Keras OCR based on deep learning and 1D CNN, and the image feature was acquired through 2D CNN. According to the identification results, the accuracy of the text recognition model was 90%. The accuracy of the comparative model and the proposed model was 91.9% and 97.6%. The accuracy, precision, recall, and F1-score of the proposed model were better than those of the comparative model in terms of statistical significance. As a result, we confirmed that the expansion of the range of feature improved the performance of the identification model.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.10a
• /
• pp.226-228
• /
• 2021

We propose a fall recognition system using the Pose Detection of Google ML kit using video data. Using the Pose detection algorithm, 33 three-dimensional feature points extracted from the body are used to recognize the fall. The algorithm that recognizes the fall by analyzing the extracted feature points uses k-NN. While passing through the normalization process in order not to be influenced in the size of the human body within the size of image and image, analyzing the relative movement of the feature points and the fall recognizes, thirteen of the thriteen test videos recognized the fall, showing an 100% success rate.