• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.6
• /
• pp.1111-1123
• /
• 2023

In recent years, as generative models have developed, research that threatens them has also been actively conducted. We propose a new membership inference attack against text-to-image model. Existing membership inference attacks on Text-to-Image models produced a single image as captions of query images. On the other hand, this paper uses personalized embedding in query images through Textual Inversion. And we propose a membership inference attack that effectively generates multiple images as a method of generating Adversarial Prompt. In addition, the membership inference attack is tested for the first time on the Stable Diffusion model, which is attracting attention among the Text-to-Image models, and achieve an accuracy of up to 1.00.

• International Journal of Computer Science & Network Security
• /
• v.22 no.7
• /
• pp.220-228
• /
• 2022

Image always carry useful information, detecting a text from scene images is imperative. The proposed work's purpose is to recognize scene text image, example boarding image kept on highways. Scene text detection on highways boarding's plays a vital role in road safety measures. At initial stage applying preprocessing techniques to the image is to sharpen and improve the features exist in the image. Likely, morphological operator were applied on images to remove the close gaps exists between objects. Here we proposed a two phase algorithm for extracting and recognizing text from scene images. In phase I text from scenery image is extracted by applying various image preprocessing techniques like blurring, erosion, tophat followed by applying thresholding, morphological gradient and by fixing kernel sizes, then canny edge detector is applied to detect the text contained in the scene images. In phase II text from scenery image recognized using MSER (Maximally Stable Extremal Region) and OCR; Proposed work aimed to detect the text contained in the scenery images from popular dataset repositories SVT, ICDAR 2003, MSRA-TD 500; these images were captured at various illumination and angles. Proposed algorithm produces higher accuracy in minimal execution time compared with state-of-the-art methodologies.

• Proceedings of the IEEK Conference
• /
• 2000.07b
• /
• pp.835-838
• /
• 2000

Popular methods for extracting a text region in video images are in general based on analysis of a whole image such as merge and split method, and comparison of two frames. Thus, they take long computing time due to the use of a whole image. Therefore, this paper suggests the faster method of extracting a text region without processing a whole image. The proposed method uses line sampling methods, FFT and neural networks in order to extract texts in real time. In general, text areas are found in the higher frequency domain, thus, can be characterized using FFT The candidate text areas can be thus found by applying the higher frequency characteristics to neural network. Therefore, the final text area is extracted by verifying the candidate areas. Experimental results show a perfect candidate extraction rate and about 92% text extraction rate. The strength of the proposed algorithm is its simplicity, real-time processing by not processing the entire image, and fast skipping of the images that do not contain a text.

• Korean Institute of Interior Design Journal
• /
• no.26
• /
• pp.88-94
• /
• 2001

This paper explores idea of the integration of text and image information in interior design. In this paper, we designed a structure of text and image information. Text information includes the information about materials and projects, and image information includes images of interior design. Material information consists of such as name and price of materials. Image information involves images of interior design that have been scanned and categorized into 15 groups according to the building regulation. Project information consists of construction brief and materials relevant to the image of interior design. The interior design information that is based on cases offers various information to designer and customers. In addition, the connection between text information and image information improves the quality of interior design by decreasing the trial and error in interior design processes. Finally, we discuss the method that integrates text and image information of interior design.

• Smart Media Journal
• /
• v.10 no.1
• /
• pp.63-69
• /
• 2021

When a person sees a sentence and understands the sentence, the person understands the sentence by reminiscent of the main word in the sentence as an image. Text-to-image is what allows computers to do this associative process. The previous deep learning-based text-to-image model extracts text features using Convolutional Neural Network (CNN)-Long Short Term Memory (LSTM) and bi-directional LSTM, and generates an image by inputting it to the GAN. The previous text-to-image model uses basic embedding in text feature extraction, and it takes a long time to train because images are generated using several modules. Therefore, in this research, we propose a method of extracting features by using the attention mechanism, which has improved performance in the natural language processing field, for sentence embedding, and generating an image by inputting the extracted features into the GAN. As a result of the experiment, the inception score was higher than that of the model used in the previous study, and when judged with the naked eye, an image that expresses the features well in the input sentence was created. In addition, even when a long sentence is input, an image that expresses the sentence well was created.

• The KIPS Transactions:PartC
• /
• v.17C no.1
• /
• pp.27-36
• /
• 2010

In Web registration pages and online polling applications, CAPTCHA(Completely Automated Public Turing Test To Tell Computers and Human Apart) is used for distinguishing human users from automated programs. Text-based CAPTCHAs have been widely used in many popular Web sites in which distorted text is used. However, because the advanced optical character recognition techniques can recognize the distorted texts, the reliability becomes low. Image-based CAPTCHAs have been proposed to improve the reliability of the text-based CAPTCHAs. However, these systems also are known as having some drawbacks. First, some image-based CAPTCHA systems with small number of image files in their image dictionary is not so reliable since attacker can recognize images by repeated executions of machine learning programs. Second, users may feel uncomfortable since they have to try CAPTCHA tests repeatedly when they fail to input a correct keyword. Third, some image-base CAPTCHAs require high communication cost since they should send several image files for one CAPTCHA. To solve these problems of image-based CAPTCHA, this paper proposes a new CAPTCHA based on both image and text. In this system, an image and keywords are integrated into one CAPTCHA image to give user a hint for the answer keyword. The proposed CAPTCHA can help users to input easily the answer keyword with the hint in the fused image. Also, the proposed system can reduce the communication costs since it uses only a fused image file for one CAPTCHA. To improve the reliability of the image-text fusion CAPTCHA, we also propose a dynamic building method of large image dictionary from gathering huge amount of images from theinternet with filtering phase for preserving the correctness of CAPTCHA images. In this paper, we proved that the proposed image-text fusion CAPTCHA provides users more convenience and high reliability than the image-based CAPTCHA through experiments.

• International Journal of Contents
• /
• v.13 no.4
• /
• pp.70-79
• /
• 2017

Images are an important element in patents and many experts use images to analyze a patent or to check differences between patents. However, there is little research on image analysis for patents partly because image processing is an advanced technology and typically patent images consist of visual parts as well as of text and numbers. This study suggests two methods for using image processing; the Scale Invariant Feature Transform(SIFT) algorithm and Optical Character Recognition(OCR). The first method which works with SIFT uses image feature points. Through feature matching, it can be applied to calculate the similarity between documents containing these images. And in the second method, OCR is used to extract text from the images. By using numbers which are extracted from an image, it is possible to extract the corresponding related text within the text passages. Subsequently, document similarity can be calculated based on the extracted text. Through comparing the suggested methods and an existing method based only on text for calculating the similarity, the feasibility is achieved. Additionally, the correlation between both the similarity measures is low which shows that they capture different aspects of the patent content.

• Journal of Information Processing Systems
• /
• v.10 no.1
• /
• pp.55-68
• /
• 2014

Automatic segmentation of foreground text from the background in degraded document images is very much essential for the smooth reading of the document content and recognition tasks by machine. In this paper, we present a novel approach to the binarization of degraded document images. The proposed method uses a new local contrast feature extracted based on the stroke width of text. First, a pre-processing method is carried out for noise removal. Text boundary detection is then performed on the image constructed from the contrast feature. Then local estimation follows to extract text from the background. Finally, a refinement procedure is applied to the binarized image as a post-processing step to improve the quality of the final results. Experiments and comparisons of extracting text from degraded handwriting and machine-printed document image against some well-known binarization algorithms demonstrate the effectiveness of the proposed method.

• Journal of Korea Multimedia Society
• /
• v.10 no.6
• /
• 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.

• Journal of Information Processing Systems
• /
• v.19 no.4
• /
• pp.427-438
• /
• 2023

Plenty of works have indicated that single image super-resolution (SISR) models relying on synthetic datasets are difficult to be applied to real scene text image super-resolution (STISR) for its more complex degradation. The up-to-date dataset for realistic STISR is called TextZoom, while the current methods trained on this dataset have not considered the effect of multi-scale features of text images. In this paper, a multi-scale and attention fusion model for realistic STISR is proposed. The multi-scale learning mechanism is introduced to acquire sophisticated feature representations of text images; The spatial and channel attentions are introduced to capture the local information and inter-channel interaction information of text images; At last, this paper designs a multi-scale residual attention module by skillfully fusing multi-scale learning and attention mechanisms. The experiments on TextZoom demonstrate that the model proposed increases scene text recognition's (ASTER) average recognition accuracy by 1.2% compared to text super-resolution network.