OBJECT DETECTION AND RECOGNITION USING YOLOV5X6 MODEL FROM REMOTE SENSING IMAGES

Abstract

The YOLOv5x6 model is fine-tuned to work with the specific task of the entity detection presented in complex and detailed RS images; the model shines by overcoming problems of detecting objects of various sizes inherent in the use of such images. Compared to the previous models, this model has avoided many shortcomings, for example, the problem with learning multi-dimension characteristics, how to balance between the recognition and the model’s complexity. Unlike other version of YOLO, the YOLOv5x6 uses new architectural improvements that lead to better performance. A specially, YOLOv5x6 is derived from the initial CA-YOLO model [16]. As for YOLOv5x6, it improved the CA-YOLO’s feature by adding a light-weight coordinate consideration unit in early layers to contain complete elements and reduce redundant info. This refinement consists in the insertion of a spatial pyramid pooling-fast which uses a tandem construction in the deeper layers [16]. In this module, the adoption of stochastic pooling procedures helps to improve the integration process of multiple scale features, as well as increases the inference speed, thus, reaching an optimal trade off between time consumption and accuracy. Moreover, YOLOv5x6 implements optimizations in anchor box mechanisms and loss functions to enhance object detection across a spectrum of sizes and scales. These optimizations ensure that the model maintains robust performance even in scenarios where objects exhibit significant variations in dimensions and spatial distribution. The achieving qualitative results support the hypothesis with recognizing YOLOv5x6 above other related versions and models. Remarkably, despite an enhancement of multiple point detection, the model has an average inference rate of 125 FPS. Combined with the high speed, this degree of precision makes YOLOv5x6 a viable method to apply in real-life situations where prompt objects’ detection within the framework of RS imagery is essential. Furthermore, the parallel research pursued in this paper seeks to address a wide range of YOLO models, extending to derivatives of the YOLO models such as YOLO V3- tiny, YOLO V4, YOLO V5s, YOLO V8s, and CA-Yolos. Stating the work unambiguously within the framework of the methodology of experimentation and comparing the achieved result with the results of previous research, the results ( directions for improving performance within the framework of these models ) are uncovered in the manuscript. Notably, the YOLOv5x6 model is revealed as one of the most effective to determine that YOLOv makes it possible to achieve high effective vi rates in different questions, including such difficult ones, as achieving mAP 95% and above with reference to near experimental object detection based on RS datasets. As a consequence, YOLOv5x6 is thus developmental enhancement on deep learning and computer vision tailored to the peculiarities of entity recognition from RS images [45]. This means it is able to handle multiple varying size as objects and at the same time being very efficient in terms of inferences than many other solutions as the results beneath show which indicates that this is a solution that is still developing and has even more potentials to be deployed in to the real world.