No abstract

No abstract

Over the last year, the correct wearing of facial masks in public is still a relevant matter in the fight against the COVID-19 pandemic. A popular approach that helps regulate the situation by global researchers is building smart systems for face mask detection. Following such spirit, this paper will contribute to the literature in two main aspects:(1) We first propose a new face mask detector model using the state-of-the-art RetinaFace for face localization in populous regions and the ResNet50V1 classifier to group the faces under 3 categories: correctly-worn, incorrectly-worn and no-masks-worn.(2) In order to select the ResNet50V1 as the backbone for the final model, we also analyzed its performance in accordance with another 3 classifiers on a face mask dataset beforehand. Performance metrics from the test phase have shown that our detector achieved the best accuracy among all the works compared, with 94, 59% on one test dataset and a less satisfactory 69.6% on another due to certain characteristics of the set. The code is available at: https://github.com/barbatoz0220/ Densely-populated-FMD.git

Purpose This study aims to solve problems of detecting copy-move images. With input images, the problem aims to: Confirm the original or forgery of the images, evaluate the performance of the detection and compare the proposed method’s effectiveness to the related ones. Design/methodology/approach This paper proposes an algorithm to identify copy-move images by matching the characteristics of objects in the same group. The method is carried out through two stages of grouping the objects and comparing objects’ features. The classification and clustering can improve processing time by skipping groups of only one object, and feature comparison on objects in the same group improves accuracy of the detection. YOLO5, the latest version of you only look once (YOLO) developed by Ultralytics LLC, and K-means are applied to classify and group the objects in the first stage. Then, modified Zernike moments (MZMs) and correlation coefficients are used for the features extraction and matching in the second stage. The Open Images V6 data set is used to train the YOLO5 model. The combination of YOLO5 and MZM makes the effectiveness of the proposed method for copy-move image detection with an average accuracy of 94.26% for images of benchmark and MICC-F600 and 95.37% for natural images. The outstanding feature of the method is that it can balance both processing time and accuracy in detecting duplicate regions on the image. Findings The problem is then solved by doing the following steps: Build a method to detect objects and compare their features to find the similarity if they are copy-move objects; use YOLO5 for the object detection and group the same category objects; ignore the group having only one object and extract the features of the other groups by MZMs; detect copy-move regions using K-means clustering; and calculate and compare the detection accuracy of the proposed method and related methods. Originality/value The main contributions of this paper include: Reduce the processing time by using YOLO5 in objects detection and K-means in clustering; improve the accuracy by using MZM to extract features and correlation coefficients to matching them; and implement and prove the effectiveness of the proposed method for three copy-move data sets: benchmark, MICC-F600 and author-built images.

The paper suggests a model based on the sharpness and blurriness to confirm the exact tampered areas from the suspicious ones which are detected from similar regions. In copy-move image detection, most research focus on comparing and finding areas with similar properties on the image. Actually, the same areas are not certainly done by copy-move manipulation, they may be the image texture. A model from the sharpness at the collage borderlines and the blurriness inside the image area is built to determine if the areas are really caused by the copy-move manipulation. The combination of feature extraction using oriented FAST and rotated BRIEF (ORB) and tampered region confirmation using a logistic regression model with 98% on accuracy proves the efficiency of the proposed methods.