| Abstract [eng] |
In this project, two different deep learning segmentation networks are analysed and evaluated by their capability to segment titanium dioxide particles in scanning electron microscopy images. In total, eight different models are set up, where each segmentation network is set up in four different ways. The models selected are U-Net, a semantic segmentation network, and Mask R-CNN, an instance segmentation network. Different network setups for U-Net consist of two different encoders, either ResNet-18 or ResNet-34, and these encoders are either pre-trained on ImageNet or not. Primary changes in the setup of the Mask R-CNN model consist of changing the limit of how many objects can be detected in a single image. A 396 image dataset is generated from 22 256x256 images by using data augmentation. Data augmentation primarily consists of spatial transformations such as rotations and flips. This generated dataset is split into 380 images for training and 16 images for validating the models. Additionally, 6 images that are not present in either the training or validation sets, are used for testing the model for the final evaluation. The best performing Mask R-CNN model was able to achieve on the validation set a mean intersection over union score of 0.5705, and a test mean intersection over union score of 0.6249. In contrast, the best-performing U-Net model was able to achieve a validation mean intersection over union score of 0.8586 and a test mean intersection over union score of 0.8766, outperforming the Mask R-CNN model. The best-performing model overall, U-Net with a pre-trained ResNet-34 backbone, was used to segment 3 different scanning electron microscopy images of titanium dioxide particles. With these segmented images, statistical analysis was performed. These three different images exhibited vastly different particle counts. It was determined that the particle area values are mainly distributed between 0 and 25 pixel area sizes, with the most spread-out area size distribution being on the image with the least amount of particles present. Further analysis showed that the Feret angles of these particles exhibit a bimodal distribution, which suggested diverse particle orientations. There was a strong correlation found among the variables that describe the compactness, roundness, circularity, area, and solidity of the particles. Welch’s T-test confirmed that the observed particles in each image show a significant difference from random distribution, hinting that the particle patterns are not the result of coincidence. All three images were categorised as having self-avoiding particles. The development of such a model can greatly help experts in the field by automating the segmentation process of the particles and allow them to focus on more prominent tasks. |
