• Journal of IKEEE
• /
• v.28 no.1
• /
• pp.65-71
• /
• 2024

This paper proposes a neural network that can reduce the number of parameters while reducing matching errors in occluded regions to increase the accuracy of depth maps in stereo matching. Stereo matching-based object recognition is utilized in many fields to more accurately recognize situations using images. When there are many objects in a complex image, an occluded area is generated due to overlap between objects and occlusion by background, thereby lowering the accuracy of the depth map. To solve this problem, existing research methods that create context information and combine it with the cost volume or RoIselect in the occluded area increase the complexity of neural networks, making it difficult to learn and expensive to implement. In this paper, we create a depthwise seperable neural network that enhances regional feature extraction before cost volume generation, reducing the number of parameters and proposing a neural network that is robust to occlusion errors. Compared to PSMNet, the proposed neural network reduced the number of parameters by 30%, improving 5.3% in color error and 3.6% in test loss.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.54 no.1
• /
• pp.96-110
• /
• 2017

Most of vehicle detection studies using the existing general lens or wide-angle lens have a blind spot in the rear detection situation, the image is vulnerable to noise and a variety of external environments. In this paper, we propose a method that is detection in harsh external environment with noise, blind spots, etc. First, using a fish-eye lens will help minimize blind spots compared to the wide-angle lens. When angle of the lens is growing because nonlinear radial distortion also increase, calibration was used after initializing and optimizing the distortion constant in order to ensure accuracy. In addition, the original image was analyzed along with calibration to remove fog and calibrate brightness and thereby enable detection even when visibility is obstructed due to light and dark adaptations from foggy situations or sudden changes in illumination. Fog removal generally takes a considerably significant amount of time to calculate. Thus in order to reduce the calculation time, remove the fog used the major fog removal algorithm Dark Channel Prior. While Gamma Correction was used to calibrate brightness, a brightness and contrast evaluation was conducted on the image in order to determine the Gamma Value needed for correction. The evaluation used only a part instead of the entirety of the image in order to reduce the time allotted to calculation. When the brightness and contrast values were calculated, those values were used to decided Gamma value and to correct the entire image. The brightness correction and fog removal were processed in parallel, and the images were registered as a single image to minimize the calculation time needed for all the processes. Then the feature extraction method HOG was used to detect the vehicle in the corrected image. As a result, it took 0.064 seconds per frame to detect the vehicle using image correction as proposed herein, which showed a 7.5% improvement in detection rate compared to the existing vehicle detection method.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.50 no.5
• /
• pp.83-93
• /
• 2013

Recently, the SURF algorithm, which is conjugated for object tracking system as part of many computer vision applications, is a well-known scale- and rotation-invariant feature detection algorithm. The SURF, due to its high computational complexity, there is essential to develop a hardware accelerator in order to be used on an IP in embedded environment. However, the SURF requires a huge local memory, causing many problems that increase the chip size and decrease the value of IP in ASIC and SoC system design. In this paper, we proposed a way to design a SURF algorithm in hardware with greatly reduced local memory by partitioning the algorithms into several Sub-IPs using external memory and a DMA. To justify validity of the proposed method, we developed an example of simplified object tracking algorithm. The execution speed of the hardware IP was about 31 frame/sec, the logic size was about 74Kgate in the 30nm technology with 81Kbytes local memory in the embedded system platform consisting of ARM Cortex-M0 processor, AMBA bus(AHB-lite and APB), DMA and a SDRAM controller. Hence, it can be used to the hardware IP of SoC Chip. If the image processing algorithm akin to SURF is applied to the method proposed in this paper, it is expected that it can implement an efficient hardware design for target application.

• Journal of KIISE:Software and Applications
• /
• v.36 no.4
• /
• pp.253-260
• /
• 2009

Many algorithms for computer vision and pattern recognition have recently been implemented on GPU (graphic processing unit) for faster computational times. However, the implementation has two problems. First, the programmer should master the fundamentals of the graphics shading languages that require the prior knowledge on computer graphics. Second, in a job that needs much cooperation between CPU and GPU, which is usual in image processing and pattern recognition contrary to the graphic area, CPU should generate raw feature data for GPU processing as much as possible to effectively utilize GPU performance. This paper proposes more quick and efficient implementation of neural networks on both GPU and multi-core CPU. We use CUDA (compute unified device architecture) that can be easily programmed due to its simple C language-like style instead of GPU to solve the first problem. Moreover, OpenMP (Open Multi-Processing) is used to concurrently process multiple data with single instruction on multi-core CPU, which results in effectively utilizing the memories of GPU. In the experiments, we implemented neural networks-based text extraction system using the proposed architecture, and the computational times showed about 15 times faster than implementation on only GPU without OpenMP.

• Journal of the Korean Society of Marine Environment & Safety
• /
• v.30 no.1
• /
• pp.97-107
• /
• 2024

Despite the rapid advancement in image recognition technology, achieving perfect digitization of tabular documents and handwritten documents still challenges. The purpose of this study is to improve the accuracy of digitizing the logbook by correcting errors by utilizing associated rules considered during logbook entries. Through this, it is expected to enhance the accuracy and reliability of data extracted from logbook through OCR programs. This model is to improve the accuracy of digitizing the logbook of the training ship "Saenuri" at the Mokpo Maritime University by correcting errors identified after Optical Character Recognition (OCR) program recognition. The model identified and corrected errors by utilizing associated rules considered during logbook entries. To evaluate the effect of model, the data before and after correction were divided by features, and comparisons were made between the same sailing number and the same feature. Using this model, approximately 10.6% of errors out of the total estimated error rate of about 11.8% were identified, and 56 out of 123 errors were corrected. A limitation of this study is that it only focuses on information from Dist.Run to Stand Course sections of the logbook, which contain navigational information. Future research will aim to correct more information from the logbook, including weather information, to overcome this limitation.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.43 no.3 s.309
• /
• pp.67-75
• /
• 2006

Facial expressions exhibit not only facial feature motions, but also subtle changes in illumination and appearance. Since it is difficult to generate realistic facial expressions by using only geometric deformations, detailed features such as textures should also be deformed to achieve more realistic expression. The existing methods such as the expression ratio image have drawbacks, in that detailed changes of complexion by lighting can not be generated properly. In this paper, we propose a nonlinear model for skin color change and a model-based synthesis method for facial expression that can apply realistic expression details under different lighting conditions. The proposed method is composed of the following three steps; automatic extraction of facial features using active appearance model and geometric deformation of expression using warping, generation of facial expression using a model for nonlinear skin color change, and synthesis of original face with generated expression using a blending ratio that is computed by the Euclidean distance transform. Experimental results show that the proposed method generate realistic facial expressions under various lighting conditions.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.43 no.3 s.309
• /
• pp.30-38
• /
• 2006

In this paper, we improved to drops recognition rate according to distance using distance and depth information with 3D from stereo face images. A monocular face image has problem to drops recognition rate by uncertainty information such as distance of an object, size, moving, rotation, and depth. Also, if image information was not acquired such as rotation, illumination, and pose change for recognition, it has a very many fault. So, we wish to solve such problem. Proposed method consists of an eyes detection algorithm, analysis a pose of face, md principal component analysis (PCA). We also convert the YCbCr space from the RGB for detect with fast face in a limited region. We create multi-layered relative intensity map in face candidate region and decide whether it is face from facial geometry. It can acquire the depth information of distance, eyes, and mouth in stereo face images. Proposed method detects face according to scale, moving, and rotation by using distance and depth. We train by using PCA the detected left face and estimated direction difference. Simulation results with face recognition rate of 95.83% (100cm) in the front and 98.3% with the pose change were obtained successfully. Therefore, proposed method can be used to obtain high recognition rate with an appropriate scaling and pose change according to the distance.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.21 no.8
• /
• pp.530-536
• /
• 2020

The use of minutiae by fingerprint readers is robust against presentation attacks, but one weakness is that the mismatch rate is high. Therefore, minutiae tend to be used with skeleton images. There have been many studies on security vulnerabilities in the characteristics of minutiae, but vulnerability studies on the skeleton are weak, so this study attempts to analyze the vulnerability of presentation attacks against the skeleton. To this end, we propose a method based on the skeleton to recover the original fingerprint using a learning algorithm. The proposed method includes a new learning model, Pix2Pix, which adds a latent vector to the existing Pix2Pix model, thereby generating a natural fingerprint. In the experimental results, the original fingerprint is restored using the proposed machine learning, and then, the restored fingerprint is the input for the fingerprint reader in order to achieve a good recognition rate. Thus, this study verifies that fingerprint readers using the skeleton are vulnerable to presentation attacks. The approach presented in this paper is expected to be useful in a variety of applications concerning fingerprint restoration, video security, and biometrics.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.20 no.6
• /
• pp.786-792
• /
• 2010

This paper presents a real-time lane detection method which can accurately find the lane-mark boundaries in complex road environment. Unlike many existing methods that pay much attention on the post-processing stage to fit lane-mark position among a great deal of outliers, the proposed method aims at removing those outliers as much as possible at feature extraction stage, so that the searching space at post-processing stage can be greatly reduced. To achieve this goal, a grid-based morphology operation is firstly used to generate the regions of interest (ROI) dynamically, in which a directional edge-linking algorithm with directional edge-gap closing is proposed to link edge-pixels into edge-links which lie in the valid directions, these directional edge-links are then grouped into pairs by checking the valid lane-mark width at certain height of the image. Finally, lane-mark colors are checked inside edge-link pairs in the YUV color space, and lane-mark types are estimated employing a Bayesian probability model. Experimental results show that the proposed method is effective in identifying lane-mark edges among heavy clutter edges in complex road environment, and the whole algorithm can achieve an accuracy rate around 92% at an average speed of 10ms/frame at the image size of $320{\times}240$.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.22 no.2
• /
• pp.127-133
• /
• 2021

A recent study of people physically falling focused on analyzing the motions of the falls using a recurrent neural network (RNN) and a deep learning approach to get good results from detecting 2D human poses from a single color image. In this paper, we investigate a detection method for estimating the position of the head and shoulder keypoints and the acceleration of positional change using the skeletal keypoints information extracted using PoseNet from an image obtained with a low-cost 2D RGB camera, increasing the accuracy of judgments about the falls. In particular, we propose a fall detection method based on the characteristics of post-fall posture in the fall motion-analysis method. A public data set was used to extract human skeletal features, and as a result of an experiment to find a feature extraction method that can achieve high classification accuracy, the proposed method showed a 99.8% success rate in detecting falls more effectively than a conventional, primitive skeletal data-use method.