Classification of Limestone Thin-Section Images Based on an Improved MobileNetV3 Incorporating RTEM and MFCM
This paper proposes an improved MobileNetV3 framework for limestone thin-section image classification that integrates a Rock-Texture Enhancement Module (RTEM) and a Multi-source Feature Complementation Module (MFCM) to effectively address complex texture variations and coupled features, achieving 91.11% accuracy while maintaining a low parameter count.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Geologists have long relied on thin slices of rock, mounted on glass and viewed under a microscope, to understand the Earth's history. These slides, known as thin sections, reveal the mineral composition, the size of the grains, and how the rock was cemented together. By studying these details, scientists can identify the specific type of limestone, a task that is crucial for finding oil and gas, understanding groundwater, and evaluating reservoirs. Traditionally, this work has been done by human experts who spend hours looking through microscopes, a process that is slow, tiring, and prone to differences in judgment from one person to another. The challenge lies in the fact that many types of limestone look incredibly similar; the boundaries between grains can be faint, and the textures can vary wildly even within the same rock type. To solve this, researchers have turned to artificial intelligence, hoping to teach computers to recognize these subtle patterns faster and more consistently than a human eye can.
A team of researchers from Northeast Petroleum University and Monash University has developed a new way to help computers identify these rock types. They focused on a specific type of artificial intelligence network called MobileNetV3, which is already known for being small and fast, making it suitable for use on devices with limited computing power. However, the standard version of this network was not quite good enough for the tricky job of distinguishing between very similar limestone textures. The researchers found that as the network processed an image, it tended to blur the fine details, such as the sharp edges of mineral grains or the delicate patterns of the rock's texture, which are exactly the clues needed to tell the rock types apart. To fix this, they redesigned the network to act more like a careful observer who never misses a small detail.
The team introduced two main improvements to the system. First, they added a special module designed to enhance the rock's texture. This module works by looking at the image in two different ways at the same time: one path focuses on finding the sharp outlines of the grains, while the other path looks for the overall arrangement and patterns of the texture. By combining these two views, the system becomes much better at seeing the boundaries and structures that define each rock type. Second, they created a method to bring together information from different levels of the network. Imagine a network that looks at an image first as a blurry overview, then as a slightly clearer picture, and finally as a detailed close-up. Usually, the network only uses the final, most detailed view to make a decision, often forgetting the useful clues it saw earlier. The new system, however, gathers the clues from the blurry, the medium, and the detailed views and blends them together. This ensures that the final decision is based on both the big picture and the tiny, critical details.
To make sure the system learned correctly, the researchers also added extra checkpoints during the training process. Instead of waiting until the very end to see if the computer was right, they asked it to make guesses at several stages along the way. If the computer made a mistake early on, it received immediate feedback to correct its path. This technique, known as deep supervision, helped the system learn much faster and more accurately. They also used a variety of tricks to make the training data more diverse, such as slightly changing the brightness, rotating the images, and mixing parts of different images together. This prevented the computer from memorizing specific pictures and forced it to learn the general rules of rock textures.
The researchers tested their new system on a dataset of 3,600 images representing six different types of limestone, including grainstone, packstone, and mudstone. The results were impressive. The improved system correctly identified the rock type in 91.11% of the cases, a significant jump from the 78.75% accuracy of the original, unmodified network. It also managed to do this while remaining extremely lightweight, using only about 2 million parameters, which is a tiny fraction of the size of other powerful image-recognition systems. In fact, it outperformed much larger and more complex networks that require far more computing power. The system was particularly good at distinguishing between the most difficult pairs of rock types, such as those with very fine grains that look almost identical to the naked eye.
This work demonstrates that it is possible to build a highly accurate rock-identification tool that is also small and efficient enough to be used in the field or on portable devices. By focusing on preserving the fine details of the rock texture and combining information from different levels of analysis, the researchers have created a model that bridges the gap between the need for speed and the need for precision. While the system still faces challenges with the most subtle variations between certain rock types, it represents a major step forward in automating geological analysis. The findings suggest that with the right architectural tweaks, artificial intelligence can become a reliable partner for geologists, helping them to decode the stories hidden within the stones of the Earth with a new level of clarity and speed.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.