• Phonetics and Speech Sciences
• /
• v.10 no.1
• /
• pp.33-38
• /
• 2018

Although deep neural network-based acoustic models have greatly improved the performance of automatic speech recognition (ASR), reverberation still degrades the performance of distant speech recognition in indoor environments. In this paper, we adopt the DenseNet, which has shown great performance results in image classification tasks, to improve the performance of reverberant speech recognition. The DenseNet enables the deep convolutional neural network (CNN) to be effectively trained by concatenating feature maps in each convolutional layer. In addition, we extend the concept of multi-resolution CNN to multi-resolution DenseNet for robust speech recognition in reverberant environments. We evaluate the performance of reverberant speech recognition on the single-channel ASR task in reverberant voice enhancement and recognition benchmark (REVERB) challenge 2014. According to the experimental results, the DenseNet-based acoustic models show better performance than do the conventional CNN-based ones, and the multi-resolution DenseNet provides additional performance improvement.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.5
• /
• pp.395-401
• /
• 2023

Sound Event Detection (SED) aims to identify not only sound category but also time interval for target sounds in an audio waveform. It is a critical technique in field of acoustic surveillance system and monitoring system. Recently, various models have introduced through Detection and Classification of Acoustic Scenes and Events (DCASE) Task 4. This paper explored how to design optimal parameters of DenseNet based model, which has led to outstanding performance in other recognition system. In experiment, DenseRNN as an SED model consists of DensNet-BC and bi-directional Gated Recurrent Units (GRU). This model is trained with Mean teacher model. With an event-based f-score, evaluation is performed depending on parameters, related to model architecture as well as model training, under the assessment protocol of DCASE task4. Experimental result shows that the performance goes up and has been saturated to near the best. Also, DenseRNN would be trained more effectively without dropout technique.

• Journal of Korea Multimedia Society
• /
• v.22 no.6
• /
• pp.665-675
• /
• 2019

In this paper, we compare and analyze the classification performance of deep learning algorithm Convolutional Neural Network(CNN) ac cording to ensemble generation and combining techniques. We used several CNN models(VGG16, VGG19, DenseNet121, DenseNet169, DenseNet201, ResNet18, ResNet34, ResNet50, ResNet101, ResNet152, GoogLeNet) to create 10 ensemble generation combinations and applied 6 combine techniques(average, weighted average, maximum, minimum, median, product) to the optimal combination. Experimental results, DenseNet169-VGG16-GoogLeNet combination in ensemble generation, and the product rule in ensemble combination showed the best performance. Based on this, it was concluded that ensemble in different models of high benchmarking scores is another way to get good results.

• Korean Journal of Remote Sensing
• /
• v.37 no.5_1
• /
• pp.1061-1070
• /
• 2021

In this manuscript, we tried to improve the performance of the FC-DenseNet by applying an attention gate for the classification of cropping areas. The attention gate module could facilitate the learning of a deep learning model and improve the performance of the model by injecting of spatial/spectral weights to each feature map. Crop classification was performed in the onion and garlic regions using a proposed deep learning model in which an attention gate was added to the skip connection part of FC-DenseNet. Training data was produced using various PlanetScope satellite imagery, and preprocessing was applied to minimize the problem of imbalanced training dataset. As a result of the crop classification, it was verified that the proposed deep learning model can more effectively classify the onion and garlic regions than existing FC-DenseNet algorithm.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.05a
• /
• pp.447-450
• /
• 2019

본 논문은 한국어 문장의 감성 분류를 위해 문장의 형태소, 음절, 자소를 입력으로 하는 합성곱층과 DenseNet 을 적용한 Text Multi-channel DenseNet 모델을 제안한다. 맞춤법 오류, 음소나 음절의 축약과 탈락, 은어나 비속어의 남용, 의태어 사용 등 문법적 규칙에 어긋나는 다양한 표현으로 인해 단어 기반 CNN 으로 추출 할 수 없는 특징들을 음절이나 자소에서 추출 할 수 있다. 한국어 감성분석에 형태소 기반 CNN 이 많이 쓰이고 있으나, 본 논문에서 제안한 Text Multi-channel DenseNet 모델은 형태소, 음절, 자소를 동시에 고려하고, DenseNet 에 정보를 밀집 전달하여 문장의 감성 분류의 정확도를 개선하였다. 네이버 영화 리뷰 데이터를 대상으로 실험한 결과 제안 모델은 85.96%의 정확도를 보여 Multi-channel CNN 에 비해 1.45% 더 정확하게 문장의 감성을 분류하였다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2018.06a
• /
• pp.272-275
• /
• 2018

본 논문에서는 기존 신경망 기반의 이미지 압축에 많이 사용되었던 신경망인 ResNet 을 대신하여 더 적은 개수의 파라미터를 사용하여 좋은 성능을 낼 수 있는 신경망 구조인 DenseNet 을 이미지 압축에 사용한다. 이미지 압축을 위해 사용되는 신경망 구조는 일반적으로 오토 인코더 구조인데, 병목 층에서 정보 손실이 상당히 많이 발생한다. 따라서 이미지 압축에서 신경망 내에서의 정보 전달은 상당히 중요하다. 기존의 논문에서는 이를 위해 이전의 정보를 그대로 뒤로 전달해주는 구조인 ResNet 을 사용하여 깊은 층에 대해서도 수렴이 잘 되는 결과를 보여주었다. 그러나 많은 수의 파라미터를 사용하는 단점을 해결하기 위해 본 논문에서는 DenseNet 을 이미지 압축에 사용하였고, 병목 층에서의 정보 손실로 인해 이미지의 고주파수 성분이 사라지는 현상을 해결하기 위해 원래 이미지와 JPEG2000 으로 압축한 이미지와의 차이를 추가 입력으로 넣어주어서 주관적인 화질을 개선하였다.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.150-153
• /
• 2022

This paper reports a technique that automatically extracts object shapes through Dense-Net, and subsequently, detects the objects using Mask R-CNN in a manufacturing site, in which workers and objects are mixed. It is based on the customized factory dataset by targeting workers, machines, tools, control boxes, and products as the objects. Mask R-CNN supports multi-object recognition as a well-known object recognition method, while Dense-Net effectively extracts a feature from multiple and overlapping objects. After immediate implementation using the two technologies, the object is naturally extracted from a still image of the manufacturing site to describe image. Afterwards, the result is planned to be used to detect workers' abnormal behavior by adding a label on the objects.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.5
• /
• pp.161-168
• /
• 2021

Cracks affect the robustness of infrastructures such as buildings, bridge, pavement, and pipelines. This paper presents an automated crack detection system which detect cracks in diverse surfaces. We first constructed the combined crack dataset, consists of multiple crack datasets in diverse domains presented in prior studies. Then, state-of-the-art deep learning models in computer vision tasks including VGG, ResNet, WideResNet, ResNeXt, DenseNet, and EfficientNet, were used to validate the performance of crack detection. We divided the combined dataset into train (80%) and test set (20%) to evaluate the employed models. DenseNet121 showed the highest accuracy at 96.20% with relatively low number of parameters compared to other models. Based on the validation procedures of the advanced deep learning models in crack detection task, we shed light on the cost-effective automated crack detection system which can be applied to different surfaces and structures with low computing resources.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.6
• /
• pp.697-702
• /
• 2019

We propose a multi-band multi-scale DenseNet with dilated convolution that separates background music signals from broadcast content. Dilated convolution can learn the multi-scale context information represented by spectrogram. In computer simulation experiments, the proposed architecture is shown to improve Signal to Distortion Ratio (SDR) by 0.15 dB and 0.27 dB in 0dB and -10 dB Signal to Noise Ratio (SNR) environments, respectively.

• The Journal of Korean Institute of Information Technology
• /
• v.17 no.8
• /
• pp.11-19
• /
• 2019

Against advanced image-related crimes, a high level of digital forensic methods is required. However, feature-based methods are difficult to respond to new device features by utilizing human-designed features, and deep learning-based methods should improve accuracy. This paper proposes a deep learning model to identify camera models based on DenseNet, the recent technology in the deep learning model field. To extract camera sensor features, a HPF feature extraction filter was applied. For camera model identification, we modified the number of hierarchical iterations and eliminated the Bottleneck layer and compression processing used to reduce computation. The proposed model was analyzed using the Dresden database and achieved an accuracy of 99.65% for 14 camera models. We achieved higher accuracy than previous studies and overcome their disadvantages with low accuracy for the same manufacturer.