全文下载： 
18.pdf
文章编号： 1672-6987（2025）02-0132-08 DOI： 10.16351/j.1672-6987.2025.02.018
程章翔a， 赵佰亭a*， 贾晓芬a，b（安徽理工大学 a. 电气与信息工程学院； b. 人工智能学院， 安徽 淮南 232001）
摘要： 井筒裂缝具有尺度多变、背景区分度小等特点，导致井筒微裂缝检测结果不理想。针对这一情况，基于主流的YOLOv8目标检测模型，提出了一种名为YOLO-DECT的井筒微裂缝检测网络。首先，融入可变形卷积模块（CN-CBS）增加并行子网，构建EMA-SPPCSCP层，增强骨干网络对多尺度特征的提取能力。然后，结合SPDConv和InceptionDWConv的优点设计了SPDIConv，替换网络中的部分卷积模块，增强模型对微裂缝的检测效果和学习效率。最后，在颈部及骨干网络末端融入EMA注意力机制，利用并行子网扩大感受野，避免背景噪声干扰。实验结果表明，提出的YOLO-DECT方法在自建数据集上的平均精度达到了83.6%，针对难以检测的微小裂缝类别检测精度达到了80.6%，相对原YOLOv8模型分别提高了3.4%和4.5%，浮点计算量下降了9.8%，模型保持了轻量化的水平，在GRDDC2022公开数据集上检测效果优于原网络，具有一定的泛用性。
关键词： 裂缝检测； 目标检测； 深度学习； 注意力机制； YOLOv8
中图分类号： TP 391 文献标志码： A
引用格式： 程章翔， 赵佰亭， 贾晓芬. 煤矿井筒微裂缝检测算法［J］. 青岛科技大学学报（自然科学版）， 2025， 46（2）： 132-139.
CHENG Zhangxiang， ZHAO Baiting， JIA Xiaofen. Coal mine shaft micro-crack detection algorithm［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition）， 2025， 46（2）： 132-139.
Coal Mine Shaft Micro-Crack Detection Algorithm
CHENG Zhangxianga， ZHAO Baitinga， JIA Xiaofena，b（a. College of Electrical and Information Engineering；b. College of Artificial Intelligence， Anhui University of Science and Technology， Huainan 232001， China）
Abstract： The characteristics of wellbore fractures are variable scale and small background differentiation， which lead to unsatisfactory detection results. To solve this problem， based on the mainstream YOLOv8 target detection model， a micro-fracture detection network named YOLO-DECT was proposed. Firstly， a deformable convolutional module （CN-CBS） is incorporated to add parallel subnets， and EMA-SPPCSCP layer is constructed to enhance the ability of the backbone network to extract multi-scale features. Then， combining the advantages of SPDConv and InceptionDWConv， SPDIConv is designed to replace some convolution modules in the network and enhance the detection effect and learning efficiency of the model for microcracks. Finally， EMA attention mechanism is integrated into the neck and the end of the backbone network， and parallel subnets are used to enlarge the receptive field and avoid the interference of background noise. The experimental results show that the average accuracy of the proposed YOLO-DECT method on the self-built data set reaches 83.6%， and the detection accuracy for the class of micro-cracks that are difficult to detect reaches 80.6%， which is 3.4% and 4.5% higher than that of the original YOLOv8 model， and the floating point computation is reduced by 9.8%. The model maintains the level of lightweight， and the detection effect on GRDDC2022 open data set is better than that of the original network， and it has certain universality.
Key words： crack detection； target detection； deep learning； attention mechanism； YOLOv8
收稿日期： 2024-07-11
基金项目： 国家自然科学基金项目（52174141）；安徽省自然科学基金项目（2108085ME158）.
作者简介： 程章翔（2000—），男，硕士研究生. * 通信联系人.
