两类YOLOv4-tiny简化网络及其裂缝检测性能比较

Comparison of Two Types YOLOv4-tiny Simplified Networks and Their Crack Detection Performance

摘要:

面向国内高大建筑物裂缝检测市场实际需求，考虑现有YOLOv4-tiny深度网络结构在树莓派等边缘设备上运行速度慢的缺点，使用去除第二层残差网络、增加一个maxpool池化层及改变最后一个route层连接的方法生成YOLOv4-lite1和YOLOv4-lite2两种新的简化版YOLOv4-tiny深度网络结构。使用从百度上搜索的裂缝图片生成裂缝检测的训练集、测试集和验证集数据，在Ubuntu16.04系统上使用Darknet深度学习框架进行了训练。同时，在树莓派4B上进行的实际测试表明，YOLOv4-lite1具有更快运行速度、检出率和稳定性。该研究创新点在于进一步精简了YOLOv4-tiny网络结构和最后一层route层的连接，从而获得两种新YOLOv4-tiny深度网络结构形式和较佳检测效果。

Abstract:

To meet the demands of crack detection market in domestic tall buildings， taking the shortcomings of the fact that the existing YOLOv4-tiny deep network structure runs slowly on such edge devices as Raspberry Pi into account， two novel simplified YOLOv4-tiny deep network structures， that is， YOLOv4-lite1 and YOLOv4-lite2 were deduced by removing the second residual network， as well as adding a maxpool layer and changing the connection of the last route layer in this paper. The training set， the test set， and the verification set data of crack detection were then generated by using the crack pictures downloaded from the internet， and the training is conducted on a 64-bit Ubuntu16.04 system utilizing the Darknet deep learning framework. At the same time， the actual tests on the RaspberryPI 4B show that the YOLOv4-lite1 structure has a faster running speed， detection rate， and stability compared to the YOLOv4-lite2 structure. Finally， the next step of this related work is pointed out. The innovation of this research lies in further simplifying the YOLOv4-tiny network structure and the connection of the last layer route layer， thus obtaining two new YOLOv4-tiny deep network structures and better detection results.