ERF-YOLO: A YOLO algorithm compatible with fewer parameters and higher accuracy
ERF-YOLO: A YOLO algorithm compatible with fewer parameters and higher accuracy
复制标题
ERF-YOLO：兼容更少参数、更高准确率的YOLO算法
DOI：
10.1016/j.imavis.2021.104317
复制
发表时间：
2021-11-05
影响因子：
4.7
通讯作者：
Zhi, Min
中科院分区：
文献类型：
--
作者：
Chai, Enhui;Ta, Lin;Zhi, Min
研究表明，理论感受野和有效感受野对目标检测结果有着重要的影响。有效感受野决定了理论感受野中不同位置的贡献。因此，本工作的主要目的是增加有效感受野面积和减少参数的数量。该方案实现了高精度、高速度的目标探测器。首先，算法需要优化激活函数，以提高特征提取的效率。其次，模型结构需要选择骨干网络，改进卷积层结构。然后，增强的网络需要增加的残余结构的数量和“Concat”，以提高特征提取性能。最后，网络需要将优化的卷积层和锚盒损失函数进行联合收割机，以提高锚盒的性能。该项目设计了一种具有更大有效感受野的YOLO算法（ERF-YOLO）。实验的训练和测试分别使用PASCAL VOC数据集和MS COCO数据集。实验结果表明，ERF-YOLO的参数接近YOLO v4的一半。在检测精度方面，ERF-YOLO优于当前的许多算法上级。(c)2021爱思唯尔有限公司版权所有。
Research shows that theoretical receptive field and effective receptive field are very important to target detection results. The effective receptive field determines the contribution of different positions in the theoretical receptive field. Therefore, the main purpose of this work is to increase the effective receptive field area and reduce the number of parameters. This idea obtains a high-precision and high-speed target detector. First, the algorithm needs to optimize the activation function to improve the efficiency of feature extraction. Second, the model structure needs to select the backbone network and improve the convolutional layer structure. Then, the enhanced network requires increasing the number of the residual structures and the "Concat" to improve feature extraction performance. Finally, the network needs to combine the optimized convolutional layer and the anchor box loss function to improve the performance of the anchor box. The project designed a YOLO algorithm (ERF-YOLO) with a larger effective receptive field. The training and testing of the experiment use PASCAL VOC data set and MS COCO data set respectively. Experimental results show that the parameter of ERF-YOLO is close to half of YOLO v4. In terms of detection accuracy, ERF-YOLO is superior to many current algorithms. (c) 2021 Elsevier B.V. All rights reserved.