• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.113-116
• /
• 2021

일반적으로 딥러닝 기반의 객체 검출(Object Detection)기법은 합성곱 신경망(Convolutional Neural Network, CNN)을 통해 입력된 영상의 특징(Feature)을 추출하여 이를 통해 객체 검출을 수행한다. 최근 자연어 처리 분야에서 획기적인 성능을 보인 트랜스포머(Transformer)가 영상 분류, 객체 검출과 같은 컴퓨터 비전 작업을 수행하는데 있어 경쟁력이 있음이 드러나고 있다. 본 논문에서는 YOLOv4-CSP의 CSP 블록을 개선한 one-stage 방식의 객체 검출 네트워크를 제안한다. 개선된 CSP 블록은 트랜스포머(Transformer)의 멀티 헤드 어텐션(Multi-Head Attention)과 CSP 형태의 공간 피라미드 풀링(Spatial Pyramid Pooling, SPP) 연산을 기반으로 네트워크의 Backbone과 Neck에서의 feature 학습을 돕는다. 본 실험은 MSCOCO test-dev2017 데이터 셋으로 평가하였으며 제안하는 네트워크는 YOLOv4-CSP의 경량화 모델인 YOLOv4s-mish에 대하여 평균 정밀도(Average Precision, AP)기준 2.7% 향상된 검출 정확도를 보인다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.372-376
• /
• 2023

본 연구에서는 비디오 하이라이트 검출 및 장면 추출을 위한 경량화된 모델인 Visual Context Learner (VCL)을 제안한다. 기존 연구에서는 매개변수가 고정된 CLIP을 비롯한 여러 피쳐 추출기에 학습 가능한 DETR과 같은 트랜스포머를 이어붙여서 학습을 한다. 하지만 본 연구는 경량화된 구조로 하이라이트 검출 성능을 개선시킬 수 있음을 보인다. 그리고 해당 형태로 장면 추출도 가능함을 보이며 장면 추출의 추가 연구 가능성을 시사한다. VCL은 매개변수가 고정된 CLIP에 학습가능한 프롬프트와 MLP로 하이라이트 검출과 장면 추출을 진행한다. 총 2,141개의 학습가능한 매개변수를 사용하여 하이라이트 검출의 HIT@1(>=Very Good) 성능을 기존 CLIP보다 2.71% 개선된 성능과 최소한의 장면 추출 성능을 보인다.

• Journal of Korea Multimedia Society
• /
• v.6 no.1
• /
• pp.67-77
• /
• 2003

Watermark detection has a crucial role in copyright protection and authentication for multimedia Because be the correlation -based algorithm which has widely been used in the watermark detection doesn't utilize the distributional characteristics of cover image to be marked, its performance is not optimum. So a new detection algorithm is proposed which is optimum for multiplicative watermark embedding. By relying on statistical decision method, the proposed method is derived according to the Bayes decision theory. Neyman Pearson criterion, and distribution of wavelet coefficients, thus Permitting to minimize the missed detection probability subject to a given false detection probability The superiority of the proposed method has been tested from a robustness perspective. The results confirm the superiority of the proposed technique over classical correlation -based method.

• The Journal of the Korea Contents Association
• /
• v.7 no.12
• /
• pp.311-321
• /
• 2007

For fast and robust face detection, this paper proposes an approach for face detection using fusion of heterogeneous template matching. First, we detect skin regions using a model of skin color which covers various illumination and races. After reducing a search space by region labelling and filtering, we apply template matching with skin color and edge to the detected regions. Finally, we detect a face by finding the best choice of template fusion. Experimental results show the proposed approach is more robust in skin color-like environments than with a single template matching and is fast by reducing a search space to face candidate regions. Also, using a global accumulator can reduce excessive space requirements of template matching.

• Journal of the Institute of Convergence Signal Processing
• /
• v.2 no.2
• /
• pp.7-12
• /
• 2001

In this paper, an estimation method of hybrid feature was proposed to detect linear feature such as the road network from SAR(synthetics aperture radar) images that include speckle noise. First we considered the mean intensity ratio or the statistical properties of locality neighboring regions to detect linear feature of road. The responses of both methods are combined to detect the entire road network. The purpose of this paper is to extract the segments of road and to mutually connect them according to the identical intensity road from the locally detected fusing images. The algorithm proposed in this paper is to define MRF(markov random field) model of the priori knowledge on the roads and applied it to energy function of interacting density points, and to detect the road networks by optimizing the energy function.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.27 no.7A
• /
• pp.695-704
• /
• 2002

Watermark detection plays a crucial role in multimedia copyright protection and has traditionally been tackled using correlation-based algorithms. However, correlation-based detection is not actually the best choice, as it does not utilize the distributional characteristics of the image being marked. Accordingly, an efficient watermark detection scheme for DWT coefficients is proposed as optimal for non-additive schemes. Based on the statistical decision theory, the proposed method is derived according to Bayes decision theory, the Neyman-Pearson criterion, and the distribution of the DWT coefficients, thereby minimizing the missed detection probability subject to a given false alarm probability. The proposed method was tested in the context of robustness, and the results confirmed the superiority of the proposed technique over conventional correlation-based detection method.

• Journal of IKEEE
• /
• v.23 no.1
• /
• pp.235-241
• /
• 2019

In this paper, we propose a new algorithm to detect rear-approaching vehicle using the frame similarity of ROI(Region of Interest) based on deep learning algorithm for use in agricultural machinery systems. Since the vehicle detection system for agricultural machinery needs to detect only a vehicle approaching from the rear. we use Faster R-CNN model that shows excellent accuracy rate in deep learning for vehicle detection. And we proposed an algorithm that uses the frame similarity for ROI using constrained conditions. Experimental results show that the proposed method has a detection rate of 99.9% and reduced the false positive values.

• Journal of the Korean Society of Marine Environment & Safety
• /
• v.28 no.6
• /
• pp.1054-1062
• /
• 2022

The population of Korea's solar salt-producing regions is rapidly aging, resulting in a decrease in the number of productive workers. In solar salt production, salt collection is the most labor-intensive operation because existing salt collection vehicles require human operators. Therefore, we intend to develop an unmanned solar salt collection vehicle to reduce manpower requirements. The unmanned solar salt collection vehicle is designed to identify the salt collection status and location in the salt plate via color detection, the color detection performance is a crucial consideration. Therefore, an image processing algorithm was developed to improve color detection performance. The algorithm generates an around-view image by using resizing, rotation, and perspective transformation of the input image, set the RoI to transform only the corresponding area to the HSV color model, and detects the color area through an AND operation. The detected color area was expanded and noise removed using morphological operations, and the area of the detection region was calculated using contour and image moment. The calculated area is compared with the set area to determine the location case of the collection vehicle within the salt plate. The performance was evaluated by comparing the calculated area of the final detected color to which the algorithm was applied and the area of the detected color in each step of the algorithm. It was confirmed that the color detection performance is improved by at least 25-99% for salt detection, at least 44-68% for red color, and an average of 7% for blue and an average of 15% for green. The proposed approach is well-suited to the operation of unmanned solar salt collection vehicles.

• Proceedings of the IEEK Conference
• /
• 2001.09a
• /
• pp.443-446
• /
• 2001

본 논문에서는 센서 융합과 배경 추정 기법을 이용하여 연속된 영상에서 물체 영역을 검출하는 기법을 제안하였다. IR/CCD각각의 카메라로부터 얻은 입력 영상을 정렬하고 융합하는 과정을 거친 후, 각 화소 단위의 배경 모델을 추정하고 시간이 지남에 따라 이를 갱신함으로써 물체 영역을 효과적으로 검출하는 기법을 제시하고 있다. 실험은 차량을 대상으로 하였고, 카메라가 움직이는 상황과 비교적 복잡한 환경에서도 좋은 결과를 얻을 수 있었다.

• Annual Conference of KIPS
• /
• 2015.04a
• /
• pp.795-797
• /
• 2015

태양광발전의 단점 중 하나인 불규칙 전력 생산문제로 인해, 장비 및 패널 결함에 실시간 대응하지 못하는 문제가 발생한다. 태양광패널 결함을 자동 검출하기 위해 기후데이터 및 패널 정보를 이용하여 신경망에 적용하고 역전과 알고리즘을 통해 학습하는 발전량 예측 및 실시간 결함 검출 모델을 제안한다.