【摘要】 针对现有目标检测器在处理无人机航拍图像上存在的小目标检测精度不高的问题,提出一种具有目标感知特征增强的改进YOLOv4(yolo only look once)航拍检测算法。通过深度级联的方式构建瓶颈连接注意力模块,将其嵌入至YOLOv4主干网络,强化对基础特征的提取;为充分有效地利用目标上下文,对原始网络中特征金字塔输出的多尺度特征进行聚合与校准细化;对于小目标定位不够精确的问题,通过闭环反馈与融合策略重新设计检测头部,增强小目标位置信息的特征响应。该方法在VisDrone航拍数据集上的实验结果表明,检测精度相比YOLOv4提高了4.24%,其中小目标的精度提升了约2%。更多还原

【Abstract】 Aiming at the problem of low detection precision of small objects in UAV aerial image processing of existing object detection methods,an improved YOLOv4(yolo only look once) aerial object detection algorithm with object-aware feature enhancement was proposed.The bottleneck connection attention module was built using the deep cascade method,and it was embedded in the YOLOv4 backbone network to strengthen the extraction of basic features.To fully and effectively utilize the object’s context information,the multi-scale features outputted by the feature pyramid in the original network were aggregated and calibrated for refinement.For the problem of inaccurate small object localization,a closed-loop feedback and fusion strategy were used to enhance the feature response of small object position information.Experimental results of the proposed method on the VisDrone aerial datasets demonstrate that the mean average precision is improved by 4.24% compared with YOLOv4,and the precision of small objects is increased by about 2%.更多还原