基于改进U-Net和双重注意力的岩石裂缝检测研究

Rock Fracture Detection Method Based on Improved U-Net and Dual Attention Mechanisms

摘要:

岩石裂隙在自然形成过程中呈现出形态曲折、边界模糊的复杂几何特性，且其图像常受到矿物纹理、光照不均等背景因素的严重干扰,针对U-Net等模型在处理此类复杂图像时易出现的特征信息丢失与拟合可靠性不足的问题，本文提出了一种融合改进U-Net与双重注意力机制的岩石裂隙检测方法。通过构建压缩金字塔结构对输入图像进行多尺度卷积，并结合空间与通道双重注意力机制，强化裂隙区域的响应权重，抑制非裂隙背景的干扰，克服特征丢失问题。进一步将双重注意力模块嵌入U-Net的编码器-解码器框架，利用Dice与Focal联合损失函数及L2正则化约束模型训练，增强对细小与模糊裂隙的识别能力，并通过尺度放缩实现像素级裂隙映射。实验结果表明，该方法对裂隙长度的检测误差基本稳定在1.5mm以内，对整体裂隙分布间距的检测误差低于10.0mm。

Abstract:

Rock fractures exhibit complex geometric characteristics such as intricate shapes and blurred boundaries during natural formation, and their images are often severely affected by background factors such as mineral texture and uneven lighting. In response to the problem of feature information loss and insufficient fitting reliability that U-Net and other models are prone to when processing such complex images, this paper proposes a rock fracture detection method that integrates improved U-Net and dual attention mechanism. By constructing a compressed pyramid structure to perform multi-scale convolution on the input image and combining spatial and channel dual attention mechanisms, the response weight of the crack area is strengthened, the interference of non crack backgrounds is suppressed, and the problem of feature loss is overcome. Further embedding the dual attention module into the encoder decoder framework of U-Net, using Dice and Focal joint loss functions and L2 regularization constraint model training to enhance the recognition ability of small and fuzzy cracks, and implementing pixel level crack mapping through scale scaling. The experimental results show that the detection error of this method for crack length is basically stable within 1.5mm, and the detection error for the overall crack distribution spacing is less than 10.0mm.