摘要： 通过分割刻蚀图像计算驱油率是目前衡量油藏驱替技术的重要手段。由于刻蚀图像边缘模糊、噪声大且水油像素分散，目前主流的图像分割网络分割精度低、分割速度慢，难以获得较好的分割结果。提出一种用于刻蚀图像分割的轻量可变形编解码网络（LDNet）。在特征编码阶段利用偏移量学习水油目标形状和边缘特征，改善网络的特征表达能力，并通过深度卷积和逐点卷积有效降低网络参数量。在特征融合阶段设计协同耦合注意力模块，将通道注意力进行分解，并分别沿2个空间方向聚合特征，将特征图编码为方向和位置感知的特征图，从而将位置信息嵌入到通道注意力中，提升模型的鲁棒性。实验结果表明，LDNet网络的分割精度为89.94%，模型大小仅为16.63×10⁶，在资源受限的设备中有效提高刻蚀图像的分割精度，降低驱油率误差，加快模型的推理速度。

关键词: 油藏驱替, 图像分割, 深度卷积神经网络, 可变形卷积, 注意力机制

Abstract: Calculating the oil displacement rate by segmenting the etching image is an essential approach for measurements in reservoir displacement technology.However, because of the blurred edges, strong noise, and scattered water and oil pixels in etched images, existing mainstream image segmentation networks, which have low segmentation accuracy and slow segmentation speed, are typically inefficient for etched image segmentation.To address this problem, A Lightweight Deformable Encoder-Decoder Network(LDNet) for etched image segmentation is proposed.First, the lightweight deformable feature encoder module not only uses offsets to learn shape and edge features for water and oil images to improve the feature representation but also decreases the number of model parameters by introducing depthwise and pointwise convolutions.Thus, the proposed Co-Coupling Attention Module(CCAM) can encode the channel attention feature maps with orientation-aware and position-aware information, improving the robustness of the network.Experiments demonstrate that the proposed network achieves improve segmentation results, with a mean dice of 89.94% for etched image segmentation, and the number of parameters is only 16.63×10⁶.The LDNet network can effectively improve the segmentation accuracy of etched images in a computer with low memory and enhance inference efficiency while minimizing the oil displacement error.

Key words: reservoir displacement, image segmentation, deep convolutional neural network, deformable convolution, attentional mechanism

中图分类号: 

• TP391