YOLO based Object Detection Enhancement based on Quadratic Convolution

Object detection remains a critical challenge with extensive real-time applications, including autonomous vehicles, medical imaging, and surveillance systems. The field has experienced significant progress, particularly with the advent of state-of-the-art detectors employing convolutional neural network architectures. Among these, the You Only Look Once (YOLO) framework has emerged as a benchmark, excelling in balancing detection accuracy and real-time performance. Nevertheless, the intrinsic linearity of conventional convolution operations constrains the network’s capacity to model complex and hierarchical data representations. In this work, we address this critical limitation by proposing a novel nonlinear operation termed quadratic convolution. Unlike standard linear convolutions, quadratic convolution involves squaring the input image elements within the convolution process, thereby augmenting the representational power of the feature maps. We incorporated this quadratic convolution into the latest YOLOv8 detector architecture to assess its effectiveness. Experiments conducted on the widely recognized MS COCO dataset indicate that our approach yields significant improvements in detection performance of the standard YOLOv8 detector. These findings underscore the potential of quadratic convolution to enhance object detection tasks, offering a promising direction for future advancements in deep learning architecture design.