信号处理2025，Vol.41Issue(5)：886-905,20.DOI:10.12466/xhcl.2025.05.009
面向降质光电图像的脑启发无人机小目标鲁棒检测方法
Brain-Inspired Robust Detection Method for Small UAV Targets in Degraded Electro-Optical Images
摘要
Abstract
In complex environments,various types of noise interference,including sensor noise,electronic interfer-ence,and weather conditions,can significantly degrade the accuracy and robustness of Unmanned Aerial Vehicle(UAV)target detection.These interferences not only deteriorate image quality but also lead to the loss or distortion of target features,thereby negatively impacting the performance of detection models.This problem is particularly pro-nounced in UAV detection tasks,where targets are often small and the background is highly complex,making detection even more challenging under noise interference.To address this issue,this study proposes a novel target detection model inspired by the multiscale orientation-selective receptive fields of the primary visual cortex(V1),termed MORF-YOLO.This model leverages the characteristics of the human visual system,employing anisotropic Gaussian kernels to simulate the receptive field mechanisms of V1 neurons,thereby extracting multi-scale and orientation-selective features from images.Integrating a V1 visual information guidance module(MORF module)into the YOLO object detection framework,MORF-YOLO enhances the representation ability of low-level features,significantly improving the adapt-ability and robustness of the model to noise interference.To validate the effectiveness of the proposed model,we con-structed a dataset with varying noise levels based on the AntiUAV2021 dataset.We compared the performance of MORF-YOLO with those of several state-of-the-art object detection methods,including YOLOv5,DiffusionDet,and DETR.Experimental results demonstrate that MORF-YOLO achieves superior detection accuracy under both noise-free and dif-ferent noise-intensity conditions.Specifically,in strong Gaussian noise scenarios(with a noise variance of 0.18),MORF-YOLO exhibits a 5%-30%improvement in detection precision(mAP@0.5)compared with other methods.Moreover,under low and medium noise conditions,MORF-YOLO shows significantly higher precision and recall rates than the compared methods.Additionally,in blurred noise and salt pepper noise scenarios,MORF-YOLO demonstrates stronger robustness,effectively reducing false detection rates and enhancing detection stability.关键词
初级视皮层/小目标检测/无人机/噪声鲁棒性/深度卷积神经网络Key words
primary visual cortex/tiny object detection/unmanned aerial vehicle/noise robustness/deep convolu-tional neural networks分类
信息技术与安全科学引用本文复制引用
李茹一,柯铭,王路斌,刘珮,高晋,王刚..面向降质光电图像的脑启发无人机小目标鲁棒检测方法[J].信号处理,2025,41(5):886-905,20.基金项目
国家自然科学青年基金(62102443) （62102443）
北京市科技新星计划(20220484097) The National Natural Science Foundation of China(62102443) （20220484097）
Beijing Nova Program(20220484097) （20220484097）
