US-12626120-B2 - Automated pixel-wise labeling of rock cuttings based on convolutional neural network-based edge detection

US12626120B2US 12626120 B2US12626120 B2US 12626120B2US-12626120-B2

Abstract

A method, workflow and system for automated classification of rock cuttings based on two tasks: (i) pixel-wise labeling of rock cutting images based on a convolutional neural network-based edge detection scheme; and (ii) training a general purpose deep learning model for classification of heterogeneous rock cutting mixtures based on the underlying texture.

Inventors

Richa Sharma

Assignees

SCHLUMBERGER TECHNOLOGY CORPORATION

Dates

Publication Date: 20260512
Application Date: 20210211

Claims (16)

1 . A computer-implemented method of classifying rock cuttings, comprising: obtaining a plurality of images depicting depicted rock cuttings distributed on a background substrate; applying one or more edge detection operators to the plurality of images to identify rock cutting boundaries within the plurality of images associated with detected edges of the depicted rock cuttings; generating binary labels for all pixels of the plurality of images based on the rock cutting boundaries, wherein generating the binary labels includes, for each image of the plurality of images: assigning a first binary label to all pixels within the rock cutting boundaries of each image of the plurality of images; and assigning a second binary label to all pixels of the background substrate of each image of the plurality of images, such that all pixels of each image of the plurality of images are assigned to only the first binary label or the second binary label; generating, without tuning the one or more edge detection operators, a plurality of training images based on modifying a color of the plurality of images and injecting a gaussian blur to the plurality of images while keeping the binary labels unchanged; causing a cutting detection machine learning model to be trained based on the plurality of training images and the first binary label assigned to all pixels and the second binary label assigned to all pixels, wherein the cutting detection machine learning model is trained to predict output binary labels for each pixel of a given image, the output binary labels indicating portions of the given image that include rock cuttings, wherein the cutting detection machine learning model includes a convolutional encoder-decoder neural network; obtaining an input image of one or more rock cuttings; and supplying the input image of the one or more rock cuttings as an input to the trained cutting detection machine learning model to automatically detect detected portions of the input image corresponding to the one or more rock cuttings in the input image and automatically generate the first binary labels indicating the portions of the input image corresponding to each of the one or more rock cuttings.
2 . The method of claim 1 , wherein obtaining the plurality of images includes selecting the plurality of images based on the depicted rock cuttings in the plurality of images not being in contact with each other and based on the plurality of images not having shadow effects.
3 . The method of claim 1 , further comprising providing the input image and the first binary labels of the detected portions of the input image to a pixel-level classification model to classify the one or more rock cuttings based on the detected portions of the input image indicated by the first binary labels.
4 . The method of claim 1 , wherein the cutting detection machine learning model is trained to automatically detect detected additional portions of the input image corresponding to dirt and fine matter in the input image and automatically assign the second binary label to all pixels of the detected additional portions corresponding to the background substrate of the input image.
5 . The method of claim 1 , wherein the plurality of images depicts downhole rock cuttings produced from a downhole operation in a downhole environment.
6 . The method of claim 1 , wherein the plurality of images depicts rock cuttings having a same lithology and color.
7 . The method of claim 1 , wherein the plurality of images depicts rock cuttings on a background of a contrasting color to a color of the rock cuttings.
8 . A system comprising: at least one processor; memory in electronic communication with the at least one processor; and instructions stored in the memory, the instructions being executable by the at least one processor to: obtain a plurality of images depicting depicted rock cuttings distributed on a background substrate; apply one or more edge detection operators to the plurality of images to identify rock cutting boundaries within the plurality of images associated with detected edges of the depicted rock cuttings; generate binary labels for all pixels of the plurality of images based on the rock cutting boundaries, wherein generating the binary labels includes, for each image of the plurality of images: assign a first binary label to all pixels within the rock cutting boundaries of each image of the plurality of images; and assign a second binary label to all pixels of the background substrate of each image of the plurality of images, such that all pixels of each image of the plurality of images are assigned to only the first binary label or the second binary label; generate, without tuning the one or more edge detection operators, a plurality of training images based on modifying a color of the plurality of images and injecting a gaussian blur to the plurality of images while keeping the binary labels unchanged; cause a cutting detection machine learning model to be trained based on the plurality of training images and the first binary label assigned to all pixels and the second binary label assigned to all pixels, wherein the cutting detection machine learning model is trained to predict output binary labels for each pixel of a given image, the output binary labels indicating portions of the given image that include rock cuttings, wherein the cutting detection machine learning model includes a convolutional encoder-decoder neural network; obtain an input image of one or more rock cuttings; and supply the input image of the one or more rock cuttings as an input to the trained cutting detection machine learning model to automatically detect detected portions of the input image corresponding to the one or more rock cuttings in the input image and automatically generate the first binary labels indicating the portions of the input image corresponding to each of the one or more rock cuttings.
9 . The system of claim 8 , wherein to obtain the plurality of images, the instructions are executable by the at least one processor to select the plurality of images based on the depicted rock cuttings in the plurality of images not being in contact with each other and based on the plurality of images not having shadow effects.
10 . The system of claim 8 , wherein the instructions are further executable by the at least one processor to provide the input image and the first binary labels of the detected portions of the input image to a pixel-level classification model to classify the one or more rock cuttings based on the detected portions of the input image indicated by the first binary labels.
11 . The system of claim 8 , wherein the cutting detection machine learning model is trained to automatically detect detected additional portions of the input image corresponding to dirt and fine matter in the input image and automatically assign the second binary label to all pixels of the detected additional portions corresponding to the background substrate of the input image.
12 . The system of claim 8 , wherein the plurality of images depicts downhole rock cuttings produced from a downhole operation in a downhole environment.
13 . The system of claim 8 , wherein the plurality of images depicts rock cuttings having a same lithology and color.
14 . The system of claim 8 , wherein the plurality of images depicts rock cuttings on a background of a contrasting color to a color of the rock cuttings.
15 . A computer-readable storage medium including instructions that, when executed by at least one processor, cause the at least one processor to: obtain a plurality of images depicting depicted rock cuttings distributed on a background substrate; apply one or more edge detection operators to the plurality of images to identify rock cutting boundaries within the plurality of images associated with detected edges of the depicted rock cuttings; generate binary labels for all pixels of the plurality of images based on the rock cutting boundaries, wherein generating the binary labels includes, for each image of the plurality of images: assign a first binary label to all pixels within the rock cutting boundaries of each image of the plurality of images; and assign a second binary label to all pixels of the background substrate of each image of the plurality of images, such that all pixels of each image of the plurality of images are assigned to only the first binary label or the second binary label; generate, without tuning the one or more edge detection operators, a plurality of training images based on modifying a color of the plurality of images and injecting a gaussian blur to the plurality of images while keeping the binary labels unchanged; cause a cutting detection machine learning model to be trained based on the plurality of training images and the first binary label assigned to all pixels and the second binary label assigned to all pixels, wherein the cutting detection machine learning model is trained to predict output binary labels for each pixel of a given image, the output binary labels indicating portions of the given image that include rock cuttings, wherein the cutting detection machine learning model includes a convolutional encoder-decoder neural network; obtain an input image of one or more rock cuttings; and supply the input image of the one or more rock cuttings as an input to the trained cutting detection machine learning model to automatically detect detected portions of the input image corresponding to the one or more rock cuttings in the input image and automatically generate the first binary labels indicating the portions of the input image corresponding to each of the one or more rock cuttings.
16 . The computer-readable storage medium of claim 15 , wherein the input image includes one or more image noise artifacts such that the one or more edge detection operators, applied to the input image, cannot detect one or more edges of the one or more rock cuttings in the input image, wherein the one or more image noise artifacts include image noise artifacts of one or more types including at least one of variations in rock cutting size distribution, variations in rock cutting texture, variations in rock cutting color, variations in rock cutting shape, variations in rock cutting spacing, or variations in rock cutting overlap, and wherein the plurality of images do not include image noise artifacts of the one or more types included in the input image.

Description

RELATED APPLICATIONS This application claims the benefit of and priority to U.S. Provisional Application Ser. No. 62/972,671, filed Feb. 11, 2020, which is incorporated by reference herein. FIELD The present disclosure relates to methods and systems that acquire images of rock cuttings and process these images for automated classification of the rock cuttings. BACKGROUND A number of sophisticated operators have been developed and tested for edge detection. Some well-known examples include: Sobel, Robert, and Prewitt operators. Each operator is designed to be sensitive to certain types of edges. The geometry of an operator governs the characteristic direction in which it is most sensitive to edges. For example, an operator can be optimized to look for horizontal, vertical, or diagonal edges. The final selection of these operators is made based on a variety of factors such as background/foreground noise, edge structure and orientation. Despite recent progress in the design, tuning, and selection of edge detection operators, significant practical challenges remain for their implementation in image classification/labeling tasks involving noisy, heterogeneous, and large-scale data. A particularly challenging application is their use in the labeling of noisy images of rock cuttings with heterogeneous cutting size distribution, texture, spacing/overlap, and background quality (shadow, brightness etc.). Furthermore, in this application a large number of images, each with multiple types of rock cuttings and lighting conditions, need to be processed at a rapid rate with relatively little human effort. Thus, the methodology and systems of the present disclosure are designed to provide automated labeling of rock cuttings in a manner that is robust and scalable. SUMMARY The present disclosure describes a method, workflow and system for automated classification of rock cuttings based on two tasks: (i) pixel-wise labeling of rock cutting images based on a convolutional neural network-based edge detection scheme; and (ii) training a general purpose deep learning model for classification of heterogeneous rock cutting mixtures based on the underlying texture. The scalability of the proposed method, workflow and system relies on automated implementation of each sub-task and their effective integration. The first task automatically generates high quality labels of rock cuttings in a highly robust, accurate, and scalable manner. The resulting pixel-wise labels are used in a second task to train a deep learning model for the classification of rock cutting images. The effectiveness of this integrated methodology and systems results from the fact that the generalization ability of deep learning models depends on the high-quality training dataset. The present disclosure describes new methods, workflows and systems that employ a substrate that supports rock cuttings during image acquisition and facilitates the processing of such images for automated classification of the rock cuttings. In embodiments, the substrate can include patterns (or a pattern of features) engraved or otherwise formed in a planar surface of material. Details of the substrate as well as workflows and systems that capture images of rock cuttings supported on the substrate are described in co-owned U.S. Provisional Patent Appl. No. 62/948,504, herein incorporated by reference in its entirety. The image-based classification or pixel wise labeling of rock cuttings can involve edge detection, which is a classical and well-studied problem in image processing. It is the process of identifying reasonably sharp discontinuities in an image. An “edge” can be defined as a discontinuity in the depth, surface, color, illumination or a combination of all these factors. Edge detection methods typically aim to achieve the following objectives: (i) maximize the probability of successfully detected edges within an image;(ii) minimize the gap between real and detected edge (accurate localization); and(iii) single detection for real edge and its shadow (unique identification). Common edge detection algorithms rely on a number of steps such as smoothing, gradient finding, non-maximum suppression, and double thresholding. In developing the methodology and systems of the present disclosure, various edge detection operators were tuned, optimized, and combined. But the results highlighted that classical edge detection methods can be extremely fragile and non-scalable. Particularly, it was found that the detected edges tend to have high-frequency noise, and further efforts to reduce this noise results in blurred and distorted edges. One possible solution to this problem could be to use operators that average enough data to discount [localized] noisy pixels. Still, in optimizing edge detection operators, one faces the fundamental tradeoff between localization accuracy and noise reduction in detected edges. The present disclosure describes a method, workflow and system for automated class