基于Yolox算法的碳酸盐岩储层溶洞“串珠状”异常反射智能检测

Intelligent detection of"bead-shaped"abnormal reflections in carbonate reservoir caves based on Yolox algorithm

识别碳酸盐岩溶洞的传统方法主要基于地震反射特征分析,对数据的要求较高,普适性不强,效率低下且存在主观因素。利用卷积神经网络(CNN)的特征提取能力识别地质构造方法的研究对象主要是盐丘、断层、地层等大尺度目标,但对于识别溶洞等小尺度构造容易出现误判。由于不同尺度的溶洞在地震剖面上呈不同的反射特征,溶洞与“串珠状”间存在一定的映射关系。因此,可首先在地震剖面上识别相对大尺度的“串珠状”,然后基于“串珠状”与溶洞的映射关系识别溶洞。为此,提出了基于Yolox的“串珠状”目标检测模型网络结构,主要包括特征提取、特征加强、解耦头三个部分。输入地震剖面后,经特征提取得到不同尺度的有效特征,然后输入特征加强网络完成多尺度的特征融合,最后由解耦头获得检测框的信息,经解码后得到“串珠状”异常反射边界的位置并输出检测框。合成地震数据测试及实际地震数据测试结果表明:(1)传统方法以振幅为基础,当“串珠状”处于强同相轴及其附近时难以识别,且高值部分无法反映“串珠状”实际范围;(2)U-Net模型识别结果反映了“串珠状”的位置,但不能获得边界坐标,易将两个很近的“串珠状”误判为一个,识别准确率不高,存在错拾或漏拾;(3)Yolox模型识别结果的检测框反映了“串珠状”的位置和实际大小,同时可得到“串珠状”的具体坐标。因此,Yolox模型识别效果优于传统方法和U-Net模型。

The traditional identification method of carbonate reservoir caves is mainly based on the analysis of seismic reflection characteristics,which has high data requirements,weak universality,and low efficiency and is affected by subjective factors.The research objects of the methods using the feature extraction capability of convolutional neural networks(CNNs)to identify geological structures are mainly large-scale structures such as salt domes,faults,and strata,but they can easily misjudge small-scale structures such as caves.Due to the different reflection characteristics of different scales of caves on the seismic profile,there is a certain mapping relationship between caves and the"bead shape."Therefore,a relatively large-scale"bead shape"can be first identified on the seismic profile,and then the cave can be identified according to the mapping relationship between"bead shape"and caves.Therefore,the network structure of the Yolox-based"bead-shaped"object detection model is proposed,which mainly includes the feature extraction module,feature enhancement module,and Decoupling Head module.Effective features of different scales are obtained by feature extraction after a seismic profile is input,and then the feature enhancement network is input for multi-scale feature fusion.Finally,the information on the detection frame is obtained by the Decoupling Head.The position of the"bead-shaped"anomalous reflection boundary is obtained after decoding,and the detection frame is output.The test results of synthetic seismic data and actual seismic data show that①the traditional method based on amplitude can hardly identify the"bead shape"when it is at and near a strong event,and the high-value part cannot reflect the actual range of the"bead shape."②The recognition results of the U-Net model reflect the position of the"bead shape"but cannot obtain the boundary coordinates,which is prone to consider two close"bead shapes"as one.Hence,it has low identification accuracy and is exposed to errors or misses.③The detection frame in the identification result of the Yolox model reflects the position and actual size of the"bead shape",and its specific coordinates can be obtained.Therefore,the identification effect of the Yolox model is better than that of traditional methods and U-Net models.