改进Pelee轻量化网络及在建筑物裂缝检测中的应用

Improved Pelee Light Network and Its Application on Crack Detection

摘要: 针对电力线绝缘子检测及高大建筑物裂缝检测等特殊应用场合，提出在边缘设备上运行深度网络程序实现智能检测的设想。综合考虑常见轻量化网络结构特点、训练难度和运行速度等特点基础上，选择Pelee为所需的轻量化深度网络。为在边缘设备上顺利运行，修改双通道全连接层结构，重新设计由主干层、残差预测层、双通道全连接层、残差预测层组成，以13 × 13为输出的深度学习网络，并在Darknet深度学习框架上设计并实现新型轻量化Pelee学习算法。建立裂缝目标检测Pascal VOC格式数据集，训练后在树莓派4B进行部署测试。结果表明，相对于YOLOv4-tiny算法，Pelee算法在权重文件大小及总计算量等方面具有明显优势，总体性能与YOLOv4-tiny相当。本文创新点在于修改双通道全连接层结构，增加残差预测模块，采用单YOLO输出头结构形式以提高小裂缝检测效果。

Abstract: In view of such special applications as power line insulator inspection with flying aircraft and crack inspection of tall buildings, the idea of running deep network program on edge equipment to realize intelligent detection is put forward in this paper. Taking the structural characteristics, training difficulty and running speed of common lightweight networks into consideration, the Pelee is selected as the required lightweight deep network. In order to run efficiently on some edge devices, the dual-channel dense layer structure was modified, and the new deep learning network consisting of stem block, residual prediction block, dual-channel dense layer and residual prediction block was redesigned with the 13 × 13 image output. Then, the Pelee lightweight learning algorithm and structure was designed on Darknet deep learning framework. After establishing dataset in Pascal VOC format for crack detection and training with Cli command, the Pelee-based crack detection deep network was deployed on RaspberryPi 4B. The results show that the Pelee algorithm has significant advantages in weight file size and total calculation in comparison to the YOLOv4-tiny network and its overall performance can be comparable to that of YOLOv4-tiny. The innovations of this paper focus on modifying structure of the dual-channel dense layer, adding an RPB module, and adopting a single YOLO output head structure to improve the detection effect of small cracks.