SEM IMAGE SEGMENTATION USING NEURAL NETWORKS WITH A MINIMAL TRAINING SET

This project adopts neural networks to study nano-scale pores in scanning electron microscope (SEM) images taken from shale samples. We built neural networks model to distinguish the objects with high accuracy and fast speed. We label the original images with different colors and analyze the RGB color ranges and assign the labels to each pixel. The output labels and the raw images are ingested as features of the training process. We train several models, predict the regions by each model, and aggregate their prediction results using a bagging approach.

The main component of the model is a U-Net variant network, which is a fully convolutional layer, consisting of a contraction path and an expansion path. There are nine units in the U-Net architecture, and each unit is composed of two convolution layers. Each unit is linked by a MaxPooling layer, a UpSampling layers, or a Concatenate layer. The copy and crop operation are used to reshape the image's dimension to ensure that the concatenation is valid.

The predicted results are evaluated with intersection over union (IOU) and accuracy metrics based on their ground truth labels. Applying our model, we achieved 97% accuracy and an average IOU of 60%. We plan to apply unsupervised learning edge detection methods to label the areas before human labeling to improve accuracy automatically. Compared to the manual segmentation method, our machine learning model predicts labels for each pixel, which the prediction results are more stable, efficient, and reliable.

The availability and accessibility of the dataset are one of the main challenges in the machine learning field. In contrast, our model does not require a large training dataset. At present, our model has obtained good results with only five training pictures. This greatly demonstrates the practicality of this model. Traditional neural network models only obtain good prediction results for particles that account for a relatively large proportion.With the adaptation of the “bagging” method that trains separated models, we solve the problem of unbalanced objects in SEM images, which widely exist in geological SEM images. Our model is deployed in a University Research Lab for SEM image segmentation and renders good performance on their custom data set.