PushPull-YOLO: A biologically inspired framework for robust object detection under image corruptions

In this study, a novel integration of PushPull-Convolutional Layers into the YOLOv11 object detection model is proposed to enhance robustness against diverse image corruptions. The PushPull-Conv layer is designed based on biological mechanisms of the primary visual cortex, where complementary push and pull kernels are utilized to improve selectivity by amplifying relevant stimuli and suppressing irrelevant noise. The initial convolutional layer of YOLOv11 is replaced by this modification, and performance is evaluated on the COCO dataset across 15 corruption types (e.g., noise, blur, weather, digital artifacts) with five severity levels. Improved robustness metrics are achieved by the PushPull-enhanced YOLOv11 compared to the baseline. Detection performance under challenging conditions, including brightness variation, motion blur, and contrast changes, is enhanced. A link is established between biologically inspired design and deep learning, positioning PushPull-YOLO as a promising solution for real-time object detection in dynamic environments, with potential extensions to segmentation and keypoint detection.