基于改进Mask R-CNN的高密度砂岩颗粒的分割识别

针对高密度颗粒密度大，数量多，形态不一，且颜色相近的情况，通过传统方法对砂岩颗粒分割难度存在检测不准和漏检的不足。想要在少量样本中获取更好的效果，变得更加困难。基于上述问题本文提出一种基于改进Mask R-CNN的DGC-Mask R-CNN检测模型，针对少量样本、高密度砂岩颗粒的分割与识别。研究中首先收集了128张超高分辨率的图片，每张图片有近200个砂岩颗粒实例，共26200个实例对象。为了使模型拥有更好的泛化能力，防止少量样本下的过拟合，使用Albu进行图像增强。用自监督预训练模型Barlow Twins来对砂岩颗粒的特征进行初步提取。在DGC-Mask R-CNN中，构建ResNet50模型作为骨干特征提取网络，在ResNet50的BottleNeck的C3，C4，C5特征卷积层中改进传统卷积方式，使用可变形卷积神经网络DCN，并添加GCB注意力机制。在上采样器的多个级联上采样模块中，结合改进的上采样算法CARAFE。实验结果表明，改进后的DGC-Mask R-CNN，使得检测与分割识别的平均精度 达到88.9%和88.8%，与传统的Mask R-CNN、Cascade-Mask R-CNN、Mask Scoring R-CNN、HybridTaskCascade相比检测精度更高。在均值平均精度 方面，与其它模型相比提升较为明显。将模型分割后得到的结果，进行砂岩颗粒的统计以及长短轴的计算，可实现对该部分砂岩颗粒的溯源，计算地壳运动导致的砂岩迁移的距离，进而评估地下油藏。

Segmentation and recognition of high-density sandstone particles based on improved Mask R-CNN

In view of the fact that the density of high-density particles is large, the number is large, the shape is different, and the color is similar, the traditional method for the segmentation of sandstone particles has the disadvantages of inaccurate detection and missing detection. To obtain better segmentation results with limited samples, it becomes more challenging. Based on the above problems, this paper proposes a DGC-Mask R-CNN detection model based on improved Mask R-CNN, aiming at the segmentation and recognition of a small number of samples and high-density sandstone particles.The study collected 128 ultra-high resolution images, each with nearly 200 sandstone particle instances, for a total of 26,200 instances, and used Albu for image augmentation to improve the model''s generalization ability and prevent overfitting with limited samples. The Barlow Twins self-supervised pre-trained model was used to extract the sandstone particle features. In Mask R-CNN, a ResNet50 model was built as the backbone feature extraction network, and the traditional convolution method was improved by using the deformable convolutional neural network (DCN) and adding the global context block (GCB) attention mechanism in the C3, C4, and C5 feature convolution layers of ResNet50''s BottleNeck. The improved upsampling algorithm CARAFE was used in multiple cascaded upsampling modules of the upsampler. The experimental results show that the improved DGC-Mask R-CNN achieves an average precision of 88.9% and 88.8% for detection and segmentation identification, respectively, which is higher than traditional Mask R-CNN, Cascade-Mask R-CNN, Mask Scoring R-CNN, and HybridTaskCascade. Compared with other models, the average accuracy of the model is significantly improved. The results obtained from the segmented model can be used to statistically analyze the sandstone particles and calculate the length and width axis to trace the source of the sandstone particles and evaluate the distance of sandstone migration caused by crustal movement in underground oil reservoirs.