Automatic image annotation system using deep learning method to analyse ambiguous images

• Published in: Periodicals of Engineering and Natural Sciences (PEN) Vol. 11; no. 2; pp. 176 - 185
• Main Author: Al-Shammary, et. al, Ali Abbas
• Format: Journal Article
• Language: English
• Published: 05.04.2023
• ISSN: 2303-4521; 2303-4521
• DOI: 10.21533/pen.v11.i2.110

Image annotation has gotten a lot of attention recently because of how quickly picture data has expanded. Together with image analysis and interpretation, image annotation, which may semantically describe imag-es, has a variety of uses in allied industries including urban planning engineering. Even without big data and image identification technologies, it is challenging to manually analyze a diverse variety of photos. The improvements to the Automated Image Annotation (AIA) label system have been the subject of several scholarly research. The authors will discuss how to use image databases and the AIA system in this essay. The proposed method extracts image features from photos using an improved VGG-19, and then uses near-by features to automatically forecast picture labels. The proposed study accounts for both correlations be-tween labels and images as well as correlations within images. The number of labels is also estimated using a label quantity prediction (LQP) model, which improves label prediction precision. The suggested method addresses automatic annotation methodologies for pixel-level images of unusual things while incorporating supervisory information via interactive spherical skins. The genuine things that were converted into metada-ta and identified as being connected to pre-existing categories were categorized by the authors using a deep learning approach called a conventional neural network (CNN) - supervised. Certain object monitoring sys-tems strive for a high item detection rate (true-positive), followed by a low availability rate (false-positive). The authors created a KD-tree based on k-nearest neighbors (KNN) to speed up annotating. In order to take into account for the collected image backdrop. The proposed method transforms the conventional two-class object detection problem into a multi-class classification problem, breaking the separated and identical dis-tribution estimations on machine learning methodologies. It is also simple to use because it only requires pixel information and ignores any other supporting elements from various color schemes. The following factors are taken into consideration while comparing the five different AIA approaches: main idea, signifi-cant contribution, computational framework, computing speed, and annotation accuracy. A set of publicly accessible photos that serve as standards for assessing AIA methods is also provided, along with a brief description of the four common assessment signs.