摘要
在加筋土桥台裂缝识别的问题中,由于裂缝过于不规则且细小裂缝较多,存在误检和识别率低等情况,为了准确地识别出加筋土桥台表面的裂缝,本研究提出一种基于YOLOv5和自适应特征融合ASFF的加筋土桥台裂缝识别算法。根据模块式土工合成材料加筋土桥台承载力模型试验,建立桥台裂缝验证数据集,利用改进的YOLOv5算法,识别桥台桥座中的裂缝区域。在模型建立时,为了对不同特征层进行上采样和下采样,将Neck结构中的特征融合算法改为ASFF自适应特征融合算法,以此来加强对细小裂缝的识别。除此之外,为避免在复杂背景的图像中出现误检的现象,本模型引入注意力机制CBAM,让模型重点关注桥台裂缝区域,抑制无用信息。经过实验表明,加筋土桥台裂缝位置识别准确率达到81.2%,平均精度89.2%,改进后的YOLOv5较原模型精度提升了5.9%,mAP@0.5提升了3.2%。结果表明基于改进后的YOLOv5加筋土桥台裂缝识别算法,提高了裂缝识别的准确度,具有较强的研究价值。
In the problem of crack identification of reinforced soil abutment,the cracks are too irregular and there are many small cracks,resulting in false detection and low recognition rate.In order to accurately identify cracks on the surface of reinforced soil abutment,a crack identification algorithm based on YOLOv5 and adaptive feature fusion ASFF is proposed in this paper.According to the bearing capacity model test of modular geosynthetic reinforced soil abutment,the data set of abutment crack verification is established,and the crack area in the abutment is identified by the improved YOLOv5 algorithm.In order to up-sample and down-sample different feature layers,the feature fusion algorithm in Neck structure was changed to ASFF adaptive feature fusion algorithm to enhance the recognition of small cracks.In addition,in order to avoid the phenomenon of false detection in images with complex background,the attention mechanism CBAM is introduced in this model to make the model focus on the crack area of the abutment and suppress useless information.Experiments show that the crack location identification accuracy of reinforced soil abutment reaches 81.2%and the average accuracy is 89.2%.Compared with the original model,the accuracy of the improved YOLOv5 is increased by 5.9% and mAP@0.5 by 3.2%.The results show that the improved YOLOv5 reinforced soil abutment crack identification algorithm improves the accuracy of crack identification and has strong research value.
作者
李莹
朱晨
Li Ying;Zhu Chen
出处
《滁州学院学报》
2024年第2期41-46,61,共7页
Journal of Chuzhou University
基金
中国地震局地震科技星火计划项目“强震动作用下台阶式加筋土挡墙破坏机理研究”(XH23067YA)
滁州城市职业学院自然科研项目“基于深度学习算法的农作物病害识别研究”(2023zkyb02)。
