基于AE-Tiny YOLOV3的小目标检测模型

Small Target Detection Model Based on AE-Tiny YOLOV3

小目标检测是现阶段目标检测领域的热点和难点问题。针对小目标检测漏检及对硬件性能要求较高的问题,对Tiny YOLOV3进行改进,提出一种适合在低性能平台上使用的小目标检测算法AE-Tiny YOLOV3。首先,使用EfficientNet-B0骨干网络替换原算法的特征提取网络;其次,在检测网络中增加一个检测分支,形成3尺度预测;最后,引入注意力机制对3个检测分支进行改进。实验结果表明,在VOC07+12数据集上,AE-Tiny YOLOV3算法满足实时检测的要求,并且鲁棒性较高,最高能将mAP值提高16.89%。将AE-Tiny YOLOV3算法应用在架空输电线路中绝缘子状态检测实例上,mAP达到了86.53%,相较于Tiny YOLOV3算法提升了15.27%,能满足对小目标绝缘子状态的实时检测。

Small target detection is a popular and difficult problem in the field of target detection in recent years.Aiming at the existing missed detection problems of small target detection and the problems with high hardware performance requirements,improves Tiny YOLOV3 and proposes a small target detection algorithm AE-Tiny YOLOV3 suitable for use on low-performance platforms.Firstly,this paper uses the EfficientNet-B0 backbone network to replace the feature extraction network of the original algorithm;secondly,add a detection branch to the detection network to make the original algorithm two-scale prediction form a three-scale prediction;finally,it introduces an attention mechanism to improve the three detection branches.ResultThe experimental results show that on the VOC07+12 dataset,the AE-Tiny YOLOV3 algorithm meets the requirements of real-time detection and has certain robustness,which can increase the mAP value by 16.89%at the highest.The AE-Tiny YOLOV3 algorithm is applied to the detection of insulator status in overhead transmission lines.The mAP reaches 86.53%,which is 15.27%higher than that of Tiny YOLOV3,which can meet the real-time detection of small target insulator status.