基于超像素与半监督的岩石图像分割与识别

岩石薄片图像的分析往往依赖于专业人员在显微镜下观察并给出鉴定结果，不但费时费力，并且受设备限制影响较大。近些年，针对于薄片图像的自动识别方法已经被提出，然而这些方法大多采用监督学习与深度学习相结合的方式，进而受限于大量人工标注，为方法的推广与应用带来了巨大困难。此外模型在不同的地层、岩性等目标应用时，其泛化性也因此受到极大限制。本文针对该问题提出了一种超像素算法SLIC与半监督自训练结合的方法，仅依靠6%的人工标注便能够实现岩石图像的自动化分割与组分识别，极大的增强该方法在实际应用中的价值。该方法首先使用超像素算法SLIC对岩石图像进行预分割，随后基于分割片的颜色特征进行粗合并，并根据最小外接矩形进行切割；切割下来的岩石组分分割图像作为后续处理的基础数据集，这里仅需要人工标注6%的岩石组分数据；随后这些数据通过一个改进的半监督自训练方法，以改进的VGG16模型作为主模型、ResNet18模型作为评判模型，不断生成高置信度的伪标签，利用迭代优化调整，将其扩展到整个数据集，最终获得一个具有较高的稳定性、准确性以及一致性的组分识别模型。通过实际数据的测试与分析，本文所提出SLIC和半监督自训练结合的方法，对6类岩石组分的识别准确率可达到96%。该方法能够在数据差异不大的条件下，帮助用户基本实现自动化的组分识别。而当数据集产生较大差异时，仅需标注少量部分样品即可实现自动的组分识别。通过理论验证与实际数据测试，本方法具有较高的泛化性和可靠性，能够在实际应用提供足够的准确性与便利性。

Semi-supervised Rock Image Segmentation and Recognition Based on Superpixel

The analysis of rock slice images often relies on professionals to observe under a microscope and give identification results, which is not only time-consuming and labor-intensive and also greatly affected by the equipment limitations. In recent years, automatic recognition methods for thin slice images have been proposed. However, most of these methods use a supervised strategy, which is limited by a large number of manual labeling and brings great difficulties to the generalization and application. In addition, when the model is applied to different reservoirs, lithology, and other targets, the generalization is greatly limited. Aiming at this problem, our paper proposes a method combining a superpixel algorithm with deep learning and semi-supervised self-training, to achieve the automatic segmentation and component identification of thin-section images only by relying on 6% of manual labels, which greatly enhances the practical application of this method. This method first uses the superpixel algorithm SLIC to pre-segment the rock image, and then performs rough merging based on the color features of the segmented slices, and cuts according to the minimum circumscribed rectangle; the cut rock component segmentation image is used as the basic data set for subsequent processing, and only 6% of the rock component data need to be manually labeled ; then these data are passed through an improved self-training method, using the improved VGG16 model as the main model and the ResNet18 model as the judgment model, and constantly generate high-confidence pseudo-labels using iteratively optimizing the tuning and extending it to the entire dataset, and finally obtain a component identification model with high stability, accuracy, and consistency. Through the testing and analysis of actual data, the method combining SLIC and semi-supervised self-training proposed in this paper can recognize 6 types of rock components with an accuracy of 96%. This method can help users basically realize automatic component identification under the condition of little data difference. When there is a large difference in the datasets, only a small number of samples can be labeled to achieve automatic component identification. Through theoretical verification and actual data testing, this method has high generalization and reliability and can provide sufficient accuracy and convenience in practical applications.