• ETRI Journal
• /
• v.46 no.3
• /
• pp.526-537
• /
• 2024

Monitoring fetal growth in utero is crucial to anomaly diagnosis. However, current computer-vision models struggle to accurately assess the key metrics (i.e., head circumference and occipitofrontal and biparietal diameters) from ultrasound images, largely owing to a lack of training data. Mitigation usually entails image augmentation (e.g., flipping, rotating, scaling, and translating). Nevertheless, the accuracy of our task remains insufficient. Hence, we offer a U-Net fetal head measurement tool that leverages a hybrid Dice and binary cross-entropy loss to compute the similarity between actual and predicted segmented regions. Ellipse-fitted two-dimensional ultrasound images acquired from the HC18 dataset are input, and their lower feature layers are reused for efficiency. During regression, a novel region of interest pooling layer extracts elliptical feature maps, and during segmentation, feature pyramids fuse field-layer data with a new scale attention method to reduce noise. Performance is measured by Dice similarity, mean pixel accuracy, and mean intersection-over-union, giving 97.90%, 99.18%, and 97.81% scores, respectively, which match or outperform the best U-Net models.

• Annual Conference of KIPS
• /
• 2024.05a
• /
• pp.317-320
• /
• 2024

최근 이미지 분류, 자율 주행 등 다양한 분야에 인공지능 기술이 접목됨에 따라 인공지능 기술을 이용한 새로운 위협이 등장하고 있다. 적대적 공격 중 단일 픽셀 공격은 이미지의 픽셀 하나를 왜곡하여 인공지능의 올바른 분류를 방해하는 공격 기법이다. 본 논문은 단일 픽셀 공격을 완화하는 이미지 처리 기법을 제안한다. 실험 결과에 따르면 제안한 방법을 적용하면 이미지의 사이즈를 27×27 로 조절하였을 때 100 개의 단일 픽셀 공격 이미지 중 94 개를 복구하였으며, 이미지의 신뢰도를 68.89% 개선하였다.

• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.166-175
• /
• 2020

Recently, application fields for processing and using digital image contents in various forms and types are rapidly increasing. Since image content is high value-added content, the intellectual property rights of this content must be protected in order to activate the production and use of the digital image content. In this paper, we propose a deep learning based watermark embedding and extraction network. The proposed method is to maximize the robustness of the watermark against malicious/non-malicious attacks while preserving the invisibility of the host image. This network consists of a preprocessing network that changes the watermark to have the same resolution as the host image, a watermark embedding network that embeds watermark data while maintaining the resolution of the host image by three-dimensionally concatenating the changed host image and the watermark information, and a watermark extraction network that reduces the resolution and extracts watermarks. This network verifies the invisibility and robustness of the proposed method by experimenting with various pixel value change attacks and geometric attacks against various watermark data and host images with various resolutions, and shows that this method is universal and practical.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.19 no.3
• /
• pp.687-692
• /
• 2018

In this paper, the development of an automotive thermal imaging camera providing image information for ADAS (Advanced Driver Assist System) and autonomous vehicles is described and an improved technique to enhance the details of the image is proposed. Thermal imaging cameras are used in various fields, such as the medical, industrial and military fields, for the purpose of temperature measurement and night vision. In automobiles, they are utilized for night vision systems. For their utilization in ADAS and autonomous vehicles, appropriate image resolution and enhanced detail are required for object recognition. In this study, a $640{\times}480$ resolution thermal imaging camera that can be applied to automobiles is developed and the BDE (Block-Range Detail Enhancement) technique is applied to improve the details of the image. In order to improve the image detail obtained in various driving environments, the block-range values between the target pixel and the surrounding 8 pixels are calculated and classified into 5 levels. Then, different factors are added or subtracted to obtain images with high utilization. The improved technique distinguishes the dark part of the image by the resulting temperature difference of 130mK and shows an improvement in the fine detail in both the bright and dark parts of the image. The developed thermal imaging camera using the improved detail enhancement technique is applied to a test vehicle and the results are presented.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.21 no.7
• /
• pp.285-291
• /
• 2020

In manufacturing, humans are being replaced with robots, but expert skills remain difficult to convert to data, making them difficult to apply to industrial robots. One method is by visual motion recognition, but physical features may be judged differently depending on the image data. This study aimed to improve the accuracy of vision methods for estimating the posture of humans. Three OpenPose vision models were applied: MPII, COCO, and COCO+foot. To identify the effects of face-covering accessories and image preprocessing on the Convolutional Neural Network (CNN) structure, the presence/non-presence of accessories, image size, and filtering were set as the parameters affecting the identification of a human's posture. For each parameter, image data were applied to the three models, and the errors between the actual and predicted values, as well as the percentage correct keypoints (PCK), were calculated. The COCO+foot model showed the lowest sensitivity to all three parameters. A <50% (from 3024×4032 to 1512×2016 pixels) reduction in image size was considered acceptable. Emboss filtering, in combination with MPII, provided the best results (reduced error of <60 pixels).

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.45 no.4
• /
• pp.10-21
• /
• 2008

The present paper propose a real-time moving object recognition and tracking method using the wavelet-based neural network and invariant moments. Candidate moving region detection phase which is the first step of the proposed method detects the candidate regions where a pixel value changes occur due to object movement based on the difference image analysis between continued two image frames. The object recognition phase which is second step of proposed method recognizes the vehicle regions from the detected candidate regions using wavelet neurual-network. From object tracking Phase which is third step the recognized vehicle regions tracks using matching methods of wavelet invariant moments bases to recognized object. To detect a moving object from image sequence the candidate regions detection phase uses an adaptive thresholding method between previous image and current image as result it was robust surroundings environmental change and moving object detections were possible. And by using wavelet features to recognize and tracking of vehicle, the proposed method decrease calculation time and not only it will be able to minimize the effect in compliance with noise of road image, vehicle recognition accuracy became improved. The result which it experiments from the image which it acquires from the general road image sequence and vehicle detection rate is 92.8%, the computing time per frame is 0.24 seconds. The proposed method can be efficiently apply to a real-time intelligence road traffic surveillance system.

• Korean Journal of Digital Imaging in Medicine
• /
• v.11 no.2
• /
• pp.115-119
• /
• 2009

I think this will be valuable reference for assuring consistency and homogeneity of clarity and managing dental radiation equipment by experimentation of dental radiation equipment permanent which based on KS C IEC 61223-3-4 standard and KS C IEC 61223-2-7. Put a dental radiation generator and experiment equipment as source and film(sensor) length within 30 em, place the step-wedge above the film(sensor). Tie up tube voltage 60 kVp, tube current 7 mA and then get an each image through CCD sensor and film by changing the exposure time as 0.12sec, 0.25sec, 0.4sec. Repeat the test 5times as a same method. Measure the concentration of each stage of film image, which gained by experiment, using photometer. And the image that gained by CCD sensor, analyze the pixel value's change by using image J, which is analyzing image program provided by NIH(National Institutes of Health). In case of film, while 0.12sec and 0.25sec show regular rising pattern of density gap as exposure time's increase, 0.4sec shows low rather than 0.12sec and 0.25sec. In case of CCD sensor density test, the result shows opposite pattern of film. This makes me think that pixels of CCD's sensor can have 0~255 value but it becomes saturation if the value is over 255. The way that getting clear reception during decreasing human's exposed radiation is one of maintaining an equipment as a best condition. So we should keeping a dental radiation equipment's condition steadily through cyclic permanent test after factor examination. Even digital equipment doesn't maintain a permanent, it can maintain a clarity by post processing of image so that hard to set it as standard of permanent test. Therefore it would be more increase the accuracy that compare a film as standard image. Thus I consider it will be an important measurement to care for dental radiation equipment and warrant homogeneity, consistency of dental image's clarity through comparing pattern which is the result from factor test against cyclic permanent test.

• Journal of Biosystems Engineering
• /
• v.26 no.1
• /
• pp.21-28
• /
• 2001

Major deficiencies of current automation scheme including various robots for bioproduction include the lack of task adaptability and real time processing, low job performance for diverse tasks, and the lack of robustness of take results, high system cost, failure of the credit from the operator, and so on. This paper proposed a scheme that could solve the current limitation of task abilities of conventional computer controlled automatic system. The proposed scheme is the man-machine hybrid automation via tele-operation which can handle various bioproduction processes. And it was classified into two categories. One category was the efficient task sharing between operator and CCM(computer controlled machine). The other was the efficient interface between operator and CCM. To realize the proposed concept, task of the object identification and extraction of 3D coordinate of an object was selected. 3D coordinate information was obtained from camera calibration using camera as a measurement device. Two stereo images were obtained by moving a camera certain distance in horizontal direction normal to focal axis and by acquiring two images at different locations. Transformation matrix for camera calibration was obtained via least square error approach using specified 6 known pairs of data points in 2D image and 3D world space. 3D world coordinate was obtained from two sets of image pixel coordinates of both camera images with calibrated transformation matrix. As an interface system between operator and CCM, a touch pad screen mounted on the monitor and remotely captured imaging system were used. Object indication was done by the operator’s finger touch to the captured image using the touch pad screen. A certain size of local image processing area was specified after the touch was made. And image processing was performed with the specified local area to extract desired features of the object. An MS Windows based interface software was developed using Visual C++6.0. The software was developed with four modules such as remote image acquisiton module, task command module, local image processing module and 3D coordinate extraction module. Proposed scheme shoed the feasibility of real time processing, robust and precise object identification, and adaptability of various job and environments though selected sample tasks.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1025-1034
• /
• 2022

In this paper, we propose a post-processing method through interpolation of hole regions that occur when extracting point clouds. When image matching is performed on stereo image data, holes occur due to occlusion and building façade area. This area may become an obstacle to the creation of additional products based on the point cloud in the future, so an effective processing technique is required. First, an initial point cloud is extracted based on the disparity map generated by applying stereo image matching. We transform the point cloud into a grid. Then a hole area is extracted due to occlusion and building façade area. By repeating the process of creating Triangulated Irregular Network (TIN) triangle in the hall area and processing the inner value of the triangle as the minimum height value of the area, it is possible to perform interpolation without awkwardness between the building and the ground surface around the building. A new point cloud is created by adding the location information corresponding to the interpolated area from the grid data as a point. To minimize the addition of unnecessary points during the interpolation process, the interpolated data to an area outside the initial point cloud area was not processed. The RGB brightness value applied to the interpolated point cloud was processed by setting the image with the closest pixel distance to the shooting center among the stereo images used for matching. It was confirmed that the shielded area generated after generating the point cloud of the target area was effectively processed through the proposed technique.

• Information Systems Review
• /
• v.25 no.1
• /
• pp.129-143
• /
• 2023

In the era of big data, the field of artificial intelligence is showing remarkable growth, and in particular, the image classification learning methods by deep learning are becoming an important area. Various studies have been actively conducted to further improve the performance of CNNs, which have been widely used in image classification, among which a representative method is the Convolutional Recurrent Neural Network (CRNN) algorithm. The CRNN algorithm consists of a combination of CNN for image classification and RNNs for recognizing time series elements. However, since the inputs used in the RNN area of CRNN are the flatten values extracted by applying the convolution and pooling technique to the image, pixel values in the same phase in the image appear in different order. And this makes it difficult to properly learn the sequence of arrangements in the image intended by the RNN. Therefore, this study aims to improve image classification performance by proposing a novel hybrid method of CNN and RNN applying the concepts of encoder and decoder. In this study, the effectiveness of the new hybrid method was verified through various experiments. This study has academic implications in that it broadens the applicability of encoder and decoder concepts, and the proposed method has advantages in terms of model learning time and infrastructure construction costs as it does not significantly increase complexity compared to conventional hybrid methods. In addition, this study has practical implications in that it presents the possibility of improving the quality of services provided in various fields that require accurate image classification.