• IEMEK Journal of Embedded Systems and Applications
• /
• v.18 no.3
• /
• pp.89-100
• /
• 2023

Recently, on-device artificial intelligence (AI) solutions using mobile devices and embedded edge devices have emerged in various fields, such as computer vision, to address network traffic burdens, low-energy operations, and security problems. Although vision transformer deep learning models have outperformed conventional convolutional neural network (CNN) models in computer vision, they require more computations and parameters than CNN models. Thus, they are not directly applicable to embedded edge devices with limited hardware resources. Many researchers have proposed various model compression methods or lightweight architectures for vision transformers; however, there are only a few studies evaluating the effects of model compression techniques of vision transformers on performance. Regarding this problem, this paper presents a performance evaluation of vision transformers on embedded platforms. We investigated the behaviors of three vision transformers: DeiT, LeViT, and MobileViT. Each model performance was evaluated by accuracy and inference time on edge devices using the ImageNet dataset. We assessed the effects of the quantization method applied to the models on latency enhancement and accuracy degradation by profiling the proportion of response time occupied by major operations. In addition, we evaluated the performance of each model on GPU and EdgeTPU-based edge devices. In our experimental results, LeViT showed the best performance in CPU-based edge devices, and DeiT-small showed the highest performance improvement in GPU-based edge devices. In addition, only MobileViT models showed performance improvement on EdgeTPU. Summarizing the analysis results through profiling, the degree of performance improvement of each vision transformer model was highly dependent on the proportion of parts that could be optimized in the target edge device. In summary, to apply vision transformers to on-device AI solutions, either proper operation composition and optimizations specific to target edge devices must be considered.

• Journal of Chest Surgery
• /
• v.11 no.1
• /
• pp.108-111
• /
• 1978

The cavernous hemangioma often occur on the skin and mucosal surfaces of the body, but are also found in many viscera, particularly the liver, spleen, pancreas and occasionally in the brain. But it is rarely encountered in the mediastinum, but when found occur predominantly in the anterior mediastinum. It can occur in any age and have no characteristic symptoms or roentgenographic findings including angiocardiography. Inspite of its histologic benignancy, it may be locally invasive and can result in rib erosion or adjacent structural compression. Usually, surgical exploration is not only the sole means of assuring a diagnosis and the only treatment. Recently, we experienced one case of cavernous hemangioma in the anterior mediastinum, which was removed surgically, being proved to be cavernous hemangioma on histologic examination. Related literatures were reviewed.

• Proceedings of the IEEK Conference
• /
• 2002.06d
• /
• pp.97-100
• /
• 2002

In this paper, we propose a watermarking technique that embeds a digital watermark into digital images for the proof of owner or copyright protection. Proposed algorithm is based on discrete wavelet transform. Zerotree structure defined by Shapiro's embedded zerotree wavelet(EZW) algorithm is used. In the proposed algorithm, a digital watermark is embedded on only significant wavelet coefficients chosen by QSWT for the robustness of the algorithm. In other words, only the values of significant wavelet coefficients are modified in accordance with the given watermark pattern. We use the relationship among neighboring coefficients when modifying chosen coefficients to keep good image quality. Visual recognizable patterns such as binary images are used as a watermark. The experimental results show that the proposed algorithm has robustness under a variety of attacks such as JPEG compression, sharpening and blurring and also show that it has a better performance in PSNR comparing with other algorithms.

• Elastomers and Composites
• /
• v.50 no.4
• /
• pp.304-313
• /
• 2015

In this study, the anionic emulsified asphalt was prepared by dispersing asphalt particles evenly into water with combination of anionic and nonionic surfactants. Effects of NaOH and $CaCl_2$ on the phase stability of the emulsified asphalt were also investigated through zeta potential value and rheology behavior; the emulsified asphalt added with NaOH and $CaCl_2$ showed higher zeta potential value than that the asphalt with addition of only anionic and nonionic surfactants. In addition, with regard to shear thinning behaviors, it was found that pH of the emulsified anionic asphalt and $Ca^{2+}$, counter ion, affected the phase stability. SBR (styrene-butadiene-rubber) latex, EPD (water dispersed Epoxy), PU (polyurethane) and RI-10S, SBS (styrene-butadiene-styrene)-based property improvement additive, were used and studied to enhance the adhesion properties with the aggregates. RI-10S, however, was found to be only compatible with the anionic emulsified asphalt; the coating rate, adhesion and compression strength were increased with the RI-10S content.

• The Transactions of the Korea Information Processing Society
• /
• v.6 no.3
• /
• pp.633-640
• /
• 1999

Reinforcement learning performs learning through interacting with trial-and-error in dynamic environment. Therefore, in dynamic environment, reinforcement learning method like Q-learning and TD(Temporal Difference)-learning are faster in learning than the conventional stochastic learning method. However, because many of the proposed reinforcement learning algorithms are given the reinforcement value only when the learning agent has reached its goal state, most of the reinforcement algorithms converge to the optimal solution too slowly. In this paper, we present COMREL(COMpressed REinforcement Learning) algorithm for finding the shortest path fast in a maze environment, select the candidate states that can guide the shortest path in compressed maze environment, and learn only the candidate states to find the shortest path. After comparing COMREL algorithm with the already existing Q-learning and Priortized Sweeping algorithm, we could see that the learning time shortened very much.

• Transactions of the Korean Society of Automotive Engineers
• /
• v.21 no.2
• /
• pp.131-135
• /
• 2013

In this study, to develop a light weight duo type aluminum LPG tank in stead of a conventional steel tank optimization technology is used. Two types of optimization method are carried out for internal compression test simulation of a LPG tank. The first is the thickness only optimization of LPG tank components. The second is the thickness and shape optimization. For the case of the thickness only optimization the weight reduction rate of an optimized tank compare to that of the initial design is 42%. Also 48% weight reduction was achieved for the case of the thickness and shape optimization.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 1999.06a
• /
• pp.57.2-62
• /
• 1999

Fractal coding is image compression techniques using one of image characteristics self-transformability. In fractal image coding, the encoding process is to select the domain block similar to a range block. The reconstructed image quality of fractal image coding depends on similitude between a range block and the selected domain block. Domain block similar to a range blocks. In fact, the error of the reconstructed image adds up the generated error in encoding process and the generated error in decoding process. But current domain block estimating function considered only the encoding error. We propose a domain block estimating function to consider not only the encoding error but also the decoding error. By computer simulation, it was verified to obtain the high quality reconstructed image.

• Journal of Information Processing Systems
• /
• v.15 no.1
• /
• pp.169-180
• /
• 2019

Transmitting visual information over a broadcasting network is not only prone to a copyright violation but also is a forgery. Authenticating such information and protecting its authorship rights call for more advanced data encoding. To this end, electronic watermarking is often adopted to embed inscriptive signature in imaging data. Most existing watermarking methods while focusing on robustness against degradation remain lacking of measurement against security loophole in which the encrypting scheme once discovered may be recreated by an unauthorized party. This could reveal the underlying signature which may potentially be replaced or forged. This paper therefore proposes a novel digital watermarking scheme in temporal-frequency domain. Unlike other typical wavelet based watermarking, the proposed scheme employed the Lorenz chaotic map to specify embedding positions. Effectively making this is not only a formidable method to decrypt but also a stronger will against deterministic attacks. Simulation report herein highlights its strength to withstand spatial and frequent adulterations, e.g., lossy compression, filtering, zooming and noise.

• Journal of fish pathology
• /
• v.34 no.1
• /
• pp.117-121
• /
• 2021

An ayu (Plecoglossus altivelis) farm in Korea recently experienced an epidemic of vertebral column deformities where about 90% of fry displayed a vertebral column that was grossly opaque along either the cranial part of the column or its entire length. Abnormal fish were lordotic, scoliotic and/or kyphotic about midway down the spine. Examination of serial sections of whole fish showed only histological lesions in the vertebral column and suggested some disturbance in the early development of the vertebral centrum. Such abnormalities included a frayed spinal or notochord sheaths with irregular thickening and compression, mal-absorbed notochord cells, thickening of around cell layer and hypercellularity on both facets of the notochord sheath. No parasites, fungi, or bacteria were detected. While this lesion has only been reported once in the past, this is the first report of histopathological findings.

• Journal of the Semiconductor & Display Technology
• /
• v.19 no.4
• /
• pp.71-76
• /
• 2020

In this paper, we propose a super-resolution method that reconstructs compressed low-resolution images into high-resolution images. We propose a CNN model with a small number of parameters, and even if quantization is applied to the proposed model, super-resolution can be implemented without deteriorating the image quality. To further improve the quality of the compressed low-resolution image, a new degradation model was proposed instead of the existing bicubic degradation model. The proposed degradation model is used only in the training process and can be applied by changing only the parameter values to the original CNN model. In the super-resolution image applying the proposed degradation model, visual artifacts caused by image compression were effectively removed. As a result, our proposed method generates higher PSNR values at compressed images and shows better visual quality, compared to conventional CNN-based SR methods.