Rock fracture detection and identification in outcrop area based on improved YOLOv7
-
摘要: 【目的】野外露头区岩石裂缝的检测对于裂缝性油气藏的地质勘测具有重要意义,但传统的图像处理算法对岩石裂缝检测效果欠佳,经典的深度学习模型网络复杂度较大、收敛与检测速度较慢且对复杂地质背景的岩石裂缝检测精度较低,因此本文提出一种改进的露头区岩石裂缝检测算法YOLOv7-PCN。【方法】首先,融入PConv(Partial Convolution)卷积模块替换主干网络的部分标准卷积,降低网络计算量提高网络检测速度;其次引入CA(Coordinate Attention)注意力机制模块,增强对裂缝关键位置特征的提取能力;最后边界框回归损失函数使用NWD(Normalized Wasserstein Distance)度量方式,优化了网络的训练收敛速度,提高了复杂环境下的分辨率较低小目标裂缝的定位检测精度。同时在数据处理方面结合数据增强方法构建露头区岩石裂缝数据集,提高网络模型的泛化能力。【结果】实验结果表明,该算法在4种岩石类别的裂缝检测上的平均精度均值达到82.5%,相比于原YOLOv7算法,提升了7.7%,同时模型参数量减少了29.6%,模型计算量节省了31.2%,模型检测速度提升了39.2%,证明了所改进算法模型的有效性。【结论】本文所提出的改进YOLOv7的岩石裂缝检测算法,综合了传统卷积神经网络模型和目前深度学习目标检测模型的优点,在拥有较少的参数量及计算量的轻量化的同时使得复杂环境下的裂缝检测结果更加准确,满足岩石裂缝检测任务的准确性和实时性要求。Abstract: [Objective]The detection of rock fractures in the outcrop area is of great significance for the geological exploration of fractured oil and gas reservoirs, but the traditional image processing algorithm is not effective for the detection of rock fractures, the classical depth learning model is characterized by high complexity, slow convergence and detection speed, and low detection accuracy for rock fractures with complex geological background, therefore, this paper presents an improved outcrop rock fracture detection algorithm YOLOv7-PCN. [Methods]Firstly, PConv (Partial Convolution) Convolution module is used to replace the standard Convolution in the backbone network, resulting in reduced computational complexity and improved detection speed. Moreover, the integration of the Coordinate Attention (CA) mechanism further improves the feature extraction capability, especially for capturing key information from critical fracture locations. Finally, the bounding box regression loss function is measured by NWD (Normalized Wasserstein Distance), which optimizes the training convergence rate of the network, the location and detection accuracy of small target fracture with lower resolution in complex environment is improved. At the same time, in data processing, A dataset of rock fractures in outcrop areas is created using data augmentation techniques to enhance the generalization capability of the network model. [Results]Experimental findings reveal an impressive average accuracy of 82.5%, surpassing the original YOLOv7 algorithm by 7.7%. Furthermore, the model's parameters are significantly reduced by 29.6%, resulting in a 31.2% reduction in computational workload. Additionally, the model's detection speed is impressively improved by 39.2%, which proves the effectiveness of the improved algorithm model. [Conclusion]The proposed enhanced YOLOv7 algorithm for rock fracture detection in this paper integrates the strengths of traditional convolutional neural network models with advanced deep learning-based target detection models, with fewer parameters and reduced computation. As a result, it achieves a higher level of accuracy in detecting rock fractures even in challenging and complex environments. The algorithm successfully meets the requirements for both accuracy and real-time performance in rock fracture detection tasks.
-
Key words:
- fracture detection /
- YOLOv7 /
- PConv /
- NWD /
- attention mechanism
点击查看大图
计量
- 文章访问数: 251
- PDF下载量: 35
- 被引次数: 0