Enhancing Remote Sensing Object Detection Through YOLOV5x6 Model
Bhasha Pydala, Bangaru Venkata Bhavana, Gorla Gamyasree, Kambam Jyotsna, Madaparthi Lohitha, V. Jyothsna
Abstract
The CA-YOLO (Coordinate Attention- YOLO) model is designed for better object detection addressing precise remote sensing pictures challenges faced by algorithms detecting multiple objects. It improves multi-scale feature extraction and tackles the compromise between detection precision and model complexity. Built on YOLOv5x6, CA-YOLO adds a thin, lightweight coordinate attention module situated in the superficial layer for efficient feature extraction and reducing redundant data. In the deeper layer, a pyramid pooling in space-fast using a tandem construction module is implemented to enhance performance without increasing model parameters. To optimize efficiency, the model reduces parameters, improves inference speed, and refines the anchor box mechanism and loss function for detecting objects of different sizes. Results demonstrate CA-YOLO outperforming the first YOLO in multiple objects detection accuracy, achieving an impressive average inference speed of 125 fps. Importantly, these improvements maintain the same parameters and complexity, making CA-YOLO an excellent choice for various applications.