基于图像区域分割和卷积神经网络的电成像缝洞表征

电成像的处理、解释大量依赖人工操作，存在缝洞表征困难等问题。人工操作不但效率低，而且还存在难以消除的人为误差。为此，提出一种基于图像区域分割和卷积神经网络的电成像图像自动识别裂缝、溶蚀孔洞的方法。该方法基于电成像数据，结合Otsu算法与平均法分割阈值，从地层背景中分离裂缝、溶蚀孔洞信息，并应用连通域像素标记法提取独立的连通域缝洞个体；然后，搭建并训练改进的LeNet-5网络模型，以多种地质构造的图像特征为标准制备训练样本集，实现缝洞特征的自动识别；最后，结合常规测井曲线，利用训练后模型的识别结果对图像分类，利用识别和提取的裂缝、溶蚀孔洞结果准确计算有效面孔率等定量评价参数。通过测试模型和实际数据的应用，验证了方法的适用性和合理性。相较于电成像的人工处理手段，该方法能够提高精度（避免人为误差）和处理速度（15s/m），训练模型针对测试集的预测准确率达97.8%，可为缝洞型储层的测井精细解释提供算法支撑。

Characterization of fractures and vugs by electrical imaging based on image region segmentation and convolutional neural network

The processing and interpretation of electrical imaging are confronted by problems including the difficulty in characterizing fractures and vugs and the dependence on manual operation. Manual ope-ration is not only inefficient but also introduces human errors which are difficult to eliminate. Therefore, this paper proposes a electrical imaging approach based on image region segmentation and the convolutional neural network to automatically identify fractures and vugs. It relies on electrical imaging data and combines with the Otsu algorithm and the average segmentation threshold to separate the fractures and vugs from the stratum background. Also, the independent fracture and vug individuals in connected domains are extracted with the connected domain pixel labeling method. Then, the automatic recognition of fractures and vugs is realized by building and training the improved LeNet-5 network model with the training sample sets based on the image features of various geological structures. Finally, according to the conventional logging curves, the recognition results of the trained model are employed to classify the images, and quantitative evaluation parameters, including effective surface porosity, are calculated accurately on the basis of identified and extracted fractures and vugs. The applicability and rationality of the proposed method are verified by the test model and actual data. At the same time, compared with the manual processing method of electrical imaging, this method can improve the accuracy (by avoiding human errors) and processing speed (15s/m), and the prediction accuracy of the trai-ning model for the test set reaches 97. 8%, providing an algorithm for the fine logging interpretation of fractured-vuggy reservoirs.