• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.24 no.1
• /
• pp.1-8
• /
• 2024

On-device AI technology, which can operate AI models at the edge devices to support real-time processing and privacy enhancement, is attracting attention. As intelligent IoT is applied to various industries, services utilizing the on-device AI technology are increasing significantly. However, general deep learning models require a lot of computational resources for inference and learning. Therefore, various lightweighting methods such as quantization and pruning have been suggested to operate deep learning models in embedded edge devices. Among the lightweighting methods, we analyze how to lightweight and apply deep learning models to edge computing devices, focusing on pruning technology in this paper. In particular, we utilize dynamic and static pruning techniques to evaluate the inference speed, accuracy, and memory usage of a lightweight AI vision model. The content analyzed in this paper can be used for intelligent video control systems or video security systems in autonomous vehicles, where real-time processing are highly required. In addition, it is expected that the content can be used more effectively in various IoT services and industries.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.17 no.6
• /
• pp.121-130
• /
• 2007

In this paper, we propose a reversible watermarking scheme in the JPEG compression domain. The reversible watermarking is useful to authenticate the content without the quality loss because it preserves the original content when embed the watermark information. In the internet, for the purpose to save the storage space and improve the efficiency of communication, digital image is usually compressed by JPEG or GIF. Therefore, it is necessary to develop a reversible watermarking in the JPEG compression domain. When the watermark is embedded, the lossless compression was used and the original image is recovered during the watermark extracting process. The test results show that PSNRs are distributed from 38dB to 42dB and the payload is from 2.5Kbits to 3.4Kbits where the QF is 75. Where the QF of the Lena image is varied from 10 to 99, the PSNR is directly proportional to the QF and the payload is around $1.6{\sim}2.8Kbits$.

• Journal of Korea Multimedia Society
• /
• v.21 no.12
• /
• pp.1396-1406
• /
• 2018

Lane detection is a key function in developing autonomous vehicle technology. In this paper, we propose a lane marker detection algorithm robust to environmental variation targeting low cost embedded computing devices. The proposed algorithm consists of two phases: initialization phase which is slow but has relatively higher accuracy; and the tracking phase which is fast and has the reliable performance in a limited condition. The initialization phase detects lane markers using a set of filters utilizing the various features of lane markers. The tracking phase uses Kalman filter to accelerate the lane marker detection processing. In a tracking phase, we measure the reliability of the detection results and switch it to initialization phase if the confidence level becomes below a threshold. By combining the initialization and tracking phases we achieved high accuracy and acceptable computing speed even under a low cost computing resources in which we cannot use the computing intensive algorithm such as deep learning approach. Experimental results show that the detection accuracy is about 95% on average and the processing speed is about 20 frames per second with Raspberry Pi 3 which is low cost device.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.44 no.6
• /
• pp.9-15
• /
• 2007

In this paper, we propose an efficient method for Processing DWT (Discrete Wavelet Transform) on GPU (Graphics Processing Unit). Since the DWT and EBCOT (embedded block coding with optimized truncation) are the most complicated submodules in JPEG2000, we design a high-performance processing framework for performing DWT using the fragment shader of GPU based on the render-to-texture (RTT) architecture. Experimental results show that the performance increases significantly, in which DWT running on modern GPU is more than 10 times faster than on modern CPU. Furthermore, by replacing the DWT part of Jasper which is the JPEG2000 reference software, the overall processing is 2$\sim$16 times faster than the original JasPer. The GPU-driven render-to-texture architecture proposed in this paper can be used in the general image and computer vision processing for high-speed processing.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.15 no.4
• /
• pp.159-166
• /
• 2020

The IBC (Intra-Body Communication) benefits from a wireless communication system for exchanging various kinds of digital information through wearable electronic devices and sensors. The IBC using the human body as the transmission channel allows wireless communication without the transmitting radio frequency waves to the air. This paper discusses the results of experiments on electrostatic coupling IBC based on FSK (Frequency Shift Keying) and 1 bit error correction. We implemented FSK communication and 1 bit error correction algorithm using the MCU boards and aluminum tape electrodes. The transmitter modulates digital data using 50% duty square wave as carrier signal and transmits data through human body. The receiver performs ADC (Analog to Digital Conversion) on carrier signal from human body. In order to figure out the frequency of carrier signal from ADC results, we applied zero-crossing algorithm which is used to detect the edge characteristic in computer vision. Experiment results shows that digital data modulated as square wave can be successfully transmitted through human body by applying the proposed architecture of a 1ch GPIO as a transmitter and 1ch ADC for as a receiver. Also, this paper proposes 1 bit error correction technique for reliable IBC. This technique performs error correction by utilizing the feature that carrier signal has 50% duty ratio. When 1 bit error correction technique is applied, the byte error rate at receiver side is improved around 3.5% compared to that not applied.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.2
• /
• pp.542-558
• /
• 2023

A digital focus index (DFI) is a value used to determine image focus in scientific apparatus and smart devices. Automatic focus (AF) is an iterative and time-consuming procedure; however, its processing time can be reduced using a general processing unit (GPU) and a multi-core processor (MCP). In this study, parallel architectures of a minimax search algorithm (MSA) are applied to two DFIs: range algorithm (RA) and image contrast (CT). The DFIs are based on a histogram; however, the parallel computation of the histogram is conventionally inefficient because of the bank conflict in shared memory. The parallel architectures of RA and CT are constructed using parallel reduction for MSA, which is performed through parallel relative rating of the image pixel pairs and halved the rating in every step. The array size is then decreased to one, and the minimax is determined at the final reduction. Kernels for the architectures are constructed using open source software to make it relatively platform independent. The kernels are tested in a hexa-core PC and an embedded device using Lenna images of various sizes based on the resolutions of industrial cameras. The performance of the kernels for the DFIs was investigated in terms of processing speed and computational acceleration; the maximum acceleration was 32.6× in the best case and the MCP exhibited a higher performance.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.10
• /
• pp.5112-5128
• /
• 2016

Tracking human body is an important problem in computer vision field. Tracking failures caused by occlusion can lead to wrong rectification of the target position. In this paper, a robust human tracking algorithm is proposed to address the problem of occlusion, rotation and improve the tracking accuracy. It is based on Tracking-Learning-Detection framework. The key auxiliary information is used in the framework which motivated by the fact that a tracking target is usually embedded in the context that provides useful information. First, face localization method is utilized to find key face location information. Second, the relative position relationship is established between the auxiliary information and the target location. With the relevant model, the key face information will get the current target position when a target has disappeared. Thus, the target can be stably tracked even when it is partially or fully occluded. Experiments are conducted in various challenging videos. In conjunction with online update, the results demonstrate that the proposed method outperforms the traditional TLD algorithm, and it has a relatively better tracking performance than other state-of-the-art methods.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.1
• /
• pp.55-62
• /
• 2019

Convolutional neural networks (CNNs) show high performance in the computer vision, but they require an enormous amount of operations, making them unsuitable for some resource- or energy-starving environments like the embedded environments. To overcome this problem, there have been much research on accelerators or pruning of CNNs. The previous pruning schemes have not considered the architecture of CNN accelerators, so the accelerators for the pruned CNNs have some inefficiency. This paper proposes a new pruning scheme, Conv-XP, which considers the architecture of CNN accelerators. In Conv-XP, the pruning is performed following the 'X' or '+' shape. The Conv-XP scheme induces a simple architecture of the CNN accelerators. The experimental results show that the Conv-XP scheme does not degrade the accuracy of CNNs, and that the accelerator area can be reduced by 12.8%.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.22 no.1
• /
• pp.1-8
• /
• 2018

Recently, the convolutional neural network shows high performance in many computer vision tasks. However, convolutional neural networks require enormous amount of operation, so it is difficult to adopt them in the embedded environments. To solve this problem, many studies are performed on the ASIC or FPGA implementation, where an efficient representation method is required. The fixed-point representation is adequate for the ASIC or FPGA implementation but causes a performance degradation. This paper proposes a separate optimization of representations for the convolutional layers and the batch normalization layers. With the proposed method, the required bit width for the convolutional layers is reduced from 16 bits to 10 bits for the ResNet-50 neural network. Since the computation amount of the convolutional layers occupies the most of the entire computation, the bit width reduction in the convolutional layers enables the efficient implementation of the convolutional neural networks.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.45 no.6
• /
• pp.133-141
• /
• 2008

As interactive multimedia have come into wide use, user interfaces such as remote controllers or classical computer mice have several limitations that cause inconvenience. We propose a vision-based pointing device to resolve this problem. We analyzed the moving image from the camera which is embedded in the pointing device and estimate the movement of the device. The pose of the cursor can be determined from this result. To process in the real time, we used the low resolution of $288{\times}208$ pixel camera and comer points of the screen were tracked using local optical flow method. The distance from screen and device was calculated from the size of screen in the image. The proposed device has simple configurations, low cost, easy use, and intuitive handhold operation like traditional mice. Moreover it shows reliable performance even in the dark condition.