Video content data have a variety of objects that could be associated with each other. Although content data contains similar objects or themes, their associations can become ambiguous. Accordingly, if associations between video content data are found in general association rules, their accuracy and confidence are low. Therefore, this study proposes the associative knowledge graph using fuzzy clustering and min-max normalization in video contents. With the use of the objects of video content, the proposed method finds clear and accurate associations between video content data and generates a knowledge graph. In the first step, the streaming video content data massively generated are collected, and objects in each image video are classified by an object detection algorithm. In the second step, normalization is executed in consideration of the different length of each video content and generate transaction data based on object frequency. In this way, it is possible to consider all the collected video content in the same condition. Additionally, it is possible to find unnecessary objects and significant objects in video content. Lastly, the degree of ambiguity is analyzed through fuzzy clustering using a probability that each object is involved in a group. Associations between fuzzy-clustered objects are extracted and an association knowledge graph is created. In this way, the accuracy and confidence of associations between different video content data are improved. As for performance, the association knowledge graph generated in the proposed method was better than a conventional association rule method in terms of the count of generated rules, support, and confidence.

As violent criminals, such as child sex offenders, tend to have high recidivism rates in modern society, there is a need to prevent such offenders from approaching socially disadvantaged and crime-prone areas, such as schools or childcare centers. Accordingly, national governments and related institutions have installed surveillance cameras and provided additional personnel to manage and monitor them via video surveillance equipment. However, naked-eye monitoring by guards and manual image processing cannot properly evaluate the video captured by surveillance cameras. To address the various problems of conventional systems that simply store and retrieve image data, a system is needed that can actively classify captured images in real-time, in addition to assisting surveillance personnel. Therefore, this paper proposes a video surveillance system based on a composable deep face recognition method. The proposed system detects the faces of criminals in real time from videos captured by a surveillance camera and notifies relevant institutions of the appearance of criminals. For real-time face detection, a down-sampled image forked from the original is used to localize unspecified faces. To improve accuracy and confidence in the recognition task, a scoring method based on face tracking is proposed. The final score combines the recognition confidence and the standard score to determine the embedding distance from the criminal face embedding data. The blind spots of surveillance personnel can be effectively addressed through early detection of criminals approaching crime-prone areas. The contributions of the paper are as follows. The proposed system can process images from surveillance cameras in real-time by using down-sampling. It can effectively identify the identity of criminals by using a face tracking ID unit and minimizes prediction reversal by solving the congested embedding problem in the feature space that may occur when performing identification matching on a large amount of face embedding DBs. Additionally, the reliability of the identification results is complemented by an identification score accumulation method. In this paper, we prototyped the proposed system and experimented with the recognition model, achieving an accuracy of 0.900 and an F-1 score of 0.943. We also experimentally confirmed that the models proposed in other studies have higher performance when using the tracked instance-level face identification method proposed in this paper. It is expected that the proposed system can be used to locate criminals and protect national facilities, and such responses can quickly prevent accidents/incidents. The dataset and code are available at https://github.com/aengoo/focusface.

This study proposes the optimization method of the associative knowledge graph using TF-IDF based ranking scores. The proposed method calculates TF-IDF weights in all documents and generates term ranking. Based on the terms with high scores from TF-IDF based ranking, optimized transactions are generated. News data are first collected through crawling and then are converted into a corpus through preprocessing. Unnecessary data are removed through preprocessing including lowercase conversion, removal of punctuation marks and stop words. In the document term matrix, words are extracted and then transactions are generated. In the data cleaning process, the Apriori algorithm is applied to generate association rules and make a knowledge graph. To optimize the generated knowledge graph, the proposed method utilizes TF-IDF based ranking scores to remove terms with low scores and recreate transactions. Based on the result, the association rule algorithm is applied to create an optimized knowledge model. The performance is evaluated in rule generation speed and usefulness of association rules. The association rule generation speed of the proposed method is about 22 seconds faster. And the lift value of the proposed method for usefulness is about 0.43 to 2.51 higher than that of each one of conventional association rule algorithms.

Reverse engineering of a stripped binary has a wide range of applications, yet it is challenging mainly due to the lack of contextually useful information within. Once debugging symbols (e.g., variable names, types, function names) are discarded, recovering such information is not technically viable with traditional approaches like static or dynamic binary analysis. We focus on a function symbol name recovery, which allows a reverse engineer to gain a quick overview of an unseen binary. The key insight is that a well-developed program labels a meaningful function name that describes its underlying semantics well. In this paper, we present AsmDepictor, the Transformer-based framework that generates a function symbol name from a set of assembly codes (i.e., machine instructions), which consists of three major components: binary code refinement, model training, and inference. To this end, we conduct systematic experiments on the effectiveness of code refinement that can enhance an overall performance. We introduce the per-layer positional embedding and Unique-softmax for AsmDepictor so that both can aid to capture a better relationship between tokens. Lastly, we devise a novel evaluation metric tailored for a short description length, the Jaccard* score. Our empirical evaluation shows that the performance of AsmDepictor by far surpasses that of the state-of-the-art models up to around 400%. The best AsmDepictor model achieves an F1 of 71.5 and Jaccard* of 75.4.

To use artificial intelligence to assist in diagnoses applications, a model to utilize quality data is required, which results in massive time and cost. In medical data, data imbalance occurs because the amount of data with lesions is less than that without lesions. To overcome this limitation, this study proposes a progressive growth of generative adversarial networks (PGGAN)-based anomaly classification on chest X-rays using weighted multi-scale similarity. An anomaly detection method is applied to learn the distribution of normal images to solve the problem of data imbalance. The use of PGGAN, which is a model that generates high-resolution images by gradually adding layers, enables to find image characteristics on a multi-scale and define the similarity between an original image and a generated image. The anomaly score is calculated by applying the weighted arithmetic mean to a resolution-by-resolution similarity. The threshold is defined after the analysis of the F1-score, and then the classification performance is evaluated. The accuracy of the proposed model was assessed using a confusion matrix and compared with that of a conventional classification model, and the efficiency was demonstrated through ablation studies. The classification accuracy of the test dataset was 0.8525. Compared to a U-net-based disease classifier with low-resolution which accuracy was 0.8410, the performance of the proposed model was 0.8507, exhibited an improvement.