| 引用本文: | 刘星,莫思特,张江,王炜康,杨世基,李鑫.轻量化模型的PeleeNet_yolov3地表裂缝识别[J].哈尔滨工业大学学报,2023,55(4):81.DOI:10.11918/202112015 |
| LIU Xing,MO Site,ZHANG Jiang,WANG Weikang,YANG Shiji,LI Xin.PeleeNet_yolov3 surface crack identification with lightweight model[J].Journal of Harbin Institute of Technology,2023,55(4):81.DOI:10.11918/202112015 |
|
| 摘要: |
| 为提高地表裂缝检测在低算力运算平台上的稳定性和检测速率,提出了一种PeleeNet与YOLOv3相结合的目标检测算法。使用PeleeNet框架代替YOLOv3的Darknet-53主体框架,以融合不同的局部特征及提高运算效率；在框架中融合特征注意力模块以提高图像中裂缝区域的显著度,并通过对感受野模块RFB卷积核的复用,增大网络的有效视野,提高小目标检测精度；在特征金字塔网络中,通过使用深度可分离卷积代替标准卷积,减少参数计算量；引入CIoU损失函数提高模型的分类与回归精度。在测试平台上应用裂缝数据进行算法验证,结果表明:AP50达到了97.68%,AP75达到了77.87%,较原始的YOLOv3分别提高8.4%和12.4%,检测速度达到了30帧/s,且模型参数大小仅为原始YOLOv3的30%；可以看出,本研究提出的PeleeNet_yolov3轻量化模型对于裂缝目标的检测效果较为明显,并且具有较小的运算量和参数量,适合应用于移动端系统,对于小体积低功耗低算力运算平台具有较大应用价值。 |
| 关键词: YOLOv3 裂缝检测 PeleeNet 深度可分离卷积 感受野模块 特征注意力模块 CIoU |
| DOI：10.11918/202112015 |
| 分类号:TP391.4 |
| 文献标识码:A |
| 基金项目:国家重点研发计划(2018YFC1505502) |
|
| PeleeNet_yolov3 surface crack identification with lightweight model |
|
LIU Xing,MO Site,ZHANG Jiang,WANG Weikang,YANG Shiji,LI Xin
|
|
(College of Electrical Engineering, Sichuan University, Chengdu 610065, China)
|
| Abstract: |
| To enhance the stability and speed of surface crack detection, a detection algorithm combining PeleeNet and YOLOv3 is proposed. PeleeNet framework is used to replace Darknet-53 framework of YOLOv3 so as to effectively integrate different local features and improve the detection rate. The feature attention module is integrated into the PeleeNet’s framework to highlight the saliency of the crack detection in the image, and through the receptive field module RFB broaden the effective field of view of the network, and increase the detection accuracy of small targets. Instead of the standard convolution, the depth separable convolution is employed to reduce the amount of parameter calculation in the feature pyramid network. Then, the CIoU loss function is introduced to strengthen the classification and regression accuracy of the model. The experimental results on the fracture data set show that AP50 and AP75 reach 97.68% and 77.87% respectively, 8.4% and 12.4% higher than the original YOLOv3. Meanwhile, the detection speed reaches 30 frames per second with the size of model parameters being only 30% of YOLOv3. As can be seen, the PeleeNet_yolov3 lightweight model proposed has produced an obvious effect on the detection of crack targets, with a small amount of calculations and parameters involved. Suitable for mobile terminal system, the study presents a great value of application especially for small volume, low power consumption and low computing power computing platforms. |
| Key words: YOLOv3 crack detection PeleeNet depth separable convolution receptive field module feature attention module CIoU |
