• 소프트웨어공학소사이어티 논문지
• /
• 제29권1호
• /
• pp.13-21
• /
• 2020

CNN 모델이 이미지 분류와 객체 탐지 등 여러 분야에 활용됨에 따라, 자율주행자동차와 같이 안전필수시스템에 사용되는 CNN 모델의 성능은 신뢰할 수 있어야 한다. 이에 CNN 모델이 다양한 환경에서도 성능을 유지하는지 평가하기 위해 배경을 변경한 이미지를 생성하는 이미지 데이터 증강 도구를 개발한다. 이미지 데이터 증강 도구에 객체가 존재하는 이미지를 입력하면, 해당 이미지로부터 객체 이미지를 추출한 후 수집한 배경 이미지 내에 객체 이미지를 합성하여 새로운 이미지를 생성한다. CNN 모델 성능 평가 방법으로 개발한 도구를 사용하여 기존 테스트 이미지로부터 새로운 테스트 이미지를 생성하고, 생성한 새로운 테스트 이미지로 CNN 모델을 평가한다. 사례 연구로 Pascal VOC2007 테스트 데이터로부터 새로운 테스트 이미지를 생성하고, 새로운 테스트 이미지로 YOLOv3 모델을 평가하였다. 그 결과 기존 테스트 이미지의 mAP 보다 새로운 테스트 이미지의 mAP가 약 0.11 더 낮아지는 것을 확인하였다.

• 한국콘텐츠학회논문지
• /
• 제16권8호
• /
• pp.592-600
• /
• 2016

AAPM CT 성능 평가용 표준 팬텀을 이용한 특수의료장비 품질관리 검사 시, 평가자의 주관적인 평가로 인한 오류를 최소화 하고자 전산화된 평가 프로그램을 이용하여 유용성을 평가하고자 한다. 평가 팬텀으로 AAPM CT Performance Phantom을 사용하였고, 기본 촬영 조건은 품질관리검사와 동일하며, 평가프로그램으로 IMAGE J를 사용하였다. 정량적인 평가로 CT 감약계수와 노이즈측정, 균일도측정, 슬라이스 두께 측정, 대조도 분해능 측정, 공간 분해능 측정의 팬텀 영상을 평가프로그램을 이용하여 영상처리를 한 후 자동추출 된 결과로 평가 하였으며, CT 감약계수, 노이즈, 균일도 측정은 영상처리를 한 영상의 표준편차가 작아 더 균일하다고 평가하였고, 슬라이스 두께 측정은 팬텀영상의 측정값과의 비율 차로 인해 평가에 어려움이 있었다. 대조도 분해능은 원통형의 지름을 6회 측정하여 지름의 평균값과 표준편차를 구해 원의 형태를 평가하였으며, 공간 분해능은 합격기준을 포함하는 원의 그룹을 자동 추출한 결과 원의 개수를 모두 추출한 결과로 나타났다. 정성적인 평가로 원본영상과 영상 처리한 영상을 육안적으로 비교 평가 하였는데 영상처리 된 영상이 우수한 결과를 나타내었다. 위의 결과 등을 바탕으로 평가자의 주관적인 판단의 오류를 최소화하기 위해서는 정량적인 평가와 정성적인 평가가 함께 이루어져야 하고 전산화된 평가프로그램을 활용한다면 보다 더 효율적인 평가가 이루어 질 것이라고 사료된다.

• 한국멀티미디어학회논문지
• /
• 제11권6호
• /
• pp.764-776
• /
• 2008

Image identification is the process of checking whether the query image is the transformed version of the specific original image or not. In this paper, image identification method based on feature composition is proposed. Used features include color distance, texture information and average pixel intensity. We extract color characteristics using color distance and texture information by Modified Generalized Symmetry Transform as well as average intensity of each pixel as features. Individual feature is quantized adaptively to be used as bins of histogram. The histogram is normalized according to data type and it is used as the signature in comparing the query image with database images. In matching part, Manhattan distance is used for measuring distance between two signatures. To evaluate the performance of the proposed method, independent test and accuracy test are achieved. In independent test, 60,433 images are used to evaluate the ability of discrimination between different images. And 4,002 original images and its 29 transformed versions are used in accuracy test, which evaluate the ability that the proposed algorithm can find the original image correctly when some transforms was applied in original image. Experiment results show that the proposed identification method has good performance in accuracy test. And the proposed method is very useful in real environment because of its high accuracy and fast matching capacity.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제9권9호
• /
• pp.3672-3684
• /
• 2015

Recently, the standardization of backward compatible JPEG image coding for high dynamic range (HDR) image has been undertaken to establish an international standard called "JPEG XT." The JPEG XT consists of two layers: the base layer and the residual layer. The base layer contains tone mapped low dynamic range (LDR) image data and the residual layer contains the error signal used to reconstruct the HDR image. This paper gives the result of a study to evaluate the overall performance of tone mapping operators (TMOs) for this standard. The evaluation is performed using five HDR image datasets and six TMOs for profiles A, B, and C of the proposed JPEG XT standard. The Tone Mapped image Quality Index (TMQI) and no reference image quality assessment (NR IQA) are used for measuring the LDR image quality. The peak signal to noise ratio (PSNR) is used to evaluate the overall compression performance of JPEG XT profiles A, B, and C. In TMQI and NR IQA measurements, TMOs using display adaptive tone mapping and adaptive logarithmic mapping each gave good results. A TMO using adaptive logarithmic mapping gave good PSNRs.

• 한국멀티미디어학회논문지
• /
• 제10권6호
• /
• pp.699-706
• /
• 2007

This article is an approach to improving the clustering of Web images by using high-level semantic features from text information relevant to Web images as well as low-level visual features of image itself. These high-level text features can be obtained from image URLs and file names, page titles, hyperlinks, and surrounding text. As a clustering algorithm, a self-organizing map (SOM) proposed by Kohonen is used. To evaluate the clustering efficiencies of SOMs, we propose a simple but effective measure indicating the accumulativeness of same class images and the perplexities of class distributions. Our approach is to advance the existing measures through defining and using new measures accumulativeness on the most superior clustering node and concentricity to evaluate clustering efficiencies of SOMs. The experimental results show that the high-level text features are more useful in SOM-based Web image clustering.

• 대한전자공학회:학술대회논문집
• /
• 대한전자공학회 1999년도 추계종합학술대회 논문집
• /
• pp.621-624
• /
• 1999

This paper describes a distance estimation method of object's motion in soccer image sequence by tracking field features. And we quantitatively evaluate the estimation accuracy We suppose that the input image sequence is taken with a camera on static axis and includes only zooming and panning transformation between frames. Adaptive template matching is adopted for non-rigid object tracking. For background compensation, feature templates selected from reference frame image are matched in following frames and the matched feature point pairs are used in computing Affine motion parameters. A perspective displacement field model is used for estimating the real distance between two position on Input Image. To quantitatively evaluate the accuracy of the estimation, we synthesized a 3 dimensional virtual stadium with graphic tools and experimented on the synthesized 2 dimensional image sequences. The experiment shows that the average of the error between the actual moving distance and the estimated distance is 1.84%.

• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2023년도 추계학술발표대회
• /
• pp.628-631
• /
• 2023

We present a new approach to evaluate the generated texts by Large Language Models (LLMs) for meme classification. Analyzing an image with embedded texts, i.e. meme, is challenging, even for existing state-of-the-art computer vision models. By leveraging large image-to-text models, we can extract image descriptions that can be used in other tasks, such as classification. In our methodology, we first generate image captions using BLIP-2 models. Using these captions, we use GPT-4 to evaluate the relationship between the caption and the meme text. The results show that OPT_6.7B provides a better rating than other LLMs, suggesting that the proposed method has a potential for meme classification.

• Biomedical Engineering Letters
• /
• 제8권4호
• /
• pp.355-364
• /
• 2018

When focusing using an ultrasonic transducer array, a main lobe is formed in the focal region of an ultrasound field, but side lobes also arise around the focal region due to the leakage. Since the side lobes cannot be completely eliminated in the focusing process, they are responsible for subsequent ultrasound image quality degradation. To improve ultrasound image quality, a signal processing strategy to reduce side lobes is definitely in demand. To this end, quantitative determination of main and side lobes is necessary. We propose a theoretically and actually error-free method of exactly discriminating and separately computing the main lobe and side lobe parts in ultrasound image by computer simulation. We refer to images constructed using the main and side lobe signals as the main and side lobe images, respectively. Since the main and side lobe images exactly represent their main and side lobe components, respectively, they can be used to evaluate ultrasound image quality. Defining the average brightness of the main and side lobe images, the conventional to side lobe image ratio, and the main to side lobe image ratio as image quality metrics, we can evaluate image characteristics in speckle images. The proposed method is also applied in assessing the performance of side lobe suppression filtering. We show that the proposed method may greatly aid in the evaluation of medical ultrasonic images using computer simulations, albeit lacking the use of actual experimental data.

• 한국조리학회지
• /
• 제11권4호
• /
• pp.30-45
• /
• 2005

This study is to explore the effects of business environment upon corporate image. Customers of foodservice establishments are selected as a population and six famous family restaurants located in Seoul were selected to evaluate the population. Accordingly, the purpose of the study is to survey satisfaction with corporate image of foodservice establishments and six famous family restaurants, to evaluate the satisfaction with corporate image, and to find a plan of raising corporate image. 240 samples were distributed and 194 samples were collected among them. 182 valid samples were selected for the research. As a result of factor analysis for business environment, internal environment factor, external environment factor, addition environment factor were extracted. There was no significant difference by sen, age, schooling, occupation, average income by month, and number of visiting foodservice establishments after examining the difference of establishment business environment by demographic characteristics. Analysing the effects of business environment upon corporate image revealed that the effects of business environment had effects upon corporate image in order of internal environment factor(.531), external environment factor(.360), and addition environment factor(.285).

• Journal of the Optical Society of Korea
• /
• 제17권5호
• /
• pp.405-409
• /
• 2013

In this paper, a unifying framework to evaluate the depth resolution of axially distributed image sensing (ADS) systems under fixed resource constraints is proposed. The proposed framework enables one to evaluate the system performance as a function of system parameters such as the number of cameras, the number of pixels, pixel size, and so on. The Monte Carlo simulations are carried out to evaluate ADS system performance as a function of system parameters. To the best of our knowledge, this is the first report on quantitative analysis of ADS systems under fixed resource constraints.