Unsupervised learning‐based recognition and extraction for intelligent automatic video retrieval

Abstract

Image queries‐based automatic recognition is a crucial element of video retrieval. Many recognition systems were proposed with different illustrations in image‐based video retrieval domains. However, intelligent automatic video retrieval recognition (IAVRR) systems development and implementation usage on video retrieval are recurring only at long intervals. Because software and hardware schemes and their updating strategies are used to necessitate automatic recognition of video retrieval image queries, to manage such functions as automated recognition of the image queries in video retrieval with a reduced degree of system memory usage, spots detection, maintenance time, identification of exact duplicate videos, etc. Simultaneously, to solve such problems, it is necessary to use video retrieval and recognition methods and algorithms that allow recognising image queries with a high degree of accuracy in a sufficient time in the conditions of interference and distortion in images. This research determination demonstrates a modified R‐ratio with Viola‐Jones (Mr. VJ) framework with the relevance of developing techniques and algorithms for automatic recognition of image queries. The proposed methods and their relevant algorithms are used to retrieve the highest real videos and ensure video retrieval's mathematical operation in the comparable protected system's operation. The protected system controls over the localisation, normalisation, parsing, segmentation and recognition are accomplished. Hence, the proposed R‐ratio analysis showed high efficiency in experiments on automatic recognition of image queries. Its application made it possible to provide a reasonably short time and high image recognition accuracy in conditions of their different locations. When using the proposed R‐ratio analysis methods with Viola‐Jones, the recognition time is about < 1000 ms. The retrieval accuracy is from 90% to 100% at the various quality of the image pixels.

Small object detection leveraging density‐aware scale adaptation

“…These well‐known detectors perform effectively on natural images (e.g., MS COCO (Lin et al, 2014) and Pascal VOC (Everingham et al, 2010)). However, in satellite‐based (or UAV‐based) surveillance (Kaliappan et al, 2022), search and rescue (Balasundaram & Krishnamoorthy, 2022), and military reconnaissance tasks where the targets are migrated to small instances, the performance of small object detection is still far from satisfactory for mainly three challenges, as shown in Figure 1: (1) small size with limited details, rendering the task of distinguishing targets from the background fairly difficult; (2) occlusion with ambiguous boundaries, leading to poor localisation; (3) variant scales with feature inconsistency, resulting in biased detection results for targets with different scales in the same scene. Therefore, small object detection, especially in densely distributed situations of remote sensing scenes, remains an open and challenging problem, requiring urgent efforts.…”