The determination of the particle size distribution (PSD) is a key test in geotechnical and civil engineering. Sieving is the standard method. It is performed all around the world thousands of times every day. Sieving is time-consuming. It is noisy and it produces dust. It also consumes water, and it retains the potential for human operators' mistakes. Recently, image-based techniques for particle sizing have become more common. Most commercial methods are based on image segmentation. The PSD is determined based on the size distribution of the segmented areas. This thesis is centered on the development of new methods based on textures and neural networks. It presents a comparative study of the performances of traditional feature extraction methods and convolutional neural networks (ConvNet). Two datasets were used: a synthetic dataset containing 53103 pairs of synthetic images produced with the discrete element code YADE, and a real dataset containing 9600 photographs of 15 materials.
The thesis has three main parts. In the first part, nine traditional feature extraction techniques were evaluated with the synthetic dataset. The features included Haralick features, histogram of oriented gradients, local entropy, local binary pattern, local configuration pattern, complete local binary pattern, fast Fourier transform, Gabor filter, and Haar discrete wavelet transform. In the second part, PSDNet, a ConvNet model, was developed, fine-tuned, and tested on the same synthetic dataset. Pretrained ConvNets were also evaluated for feature extraction and transfer learning. The pretrained ConvNet that were tested in this research include AlexNet (Krizhevsky et al., 2012), SqueezeNet (Iandola et al., 2016) , GoogLeNet (Szegedy et al., 2015), InceptionV3 (Szegedy, Vanhoucke, et al., 2016), DenseNet201 (Huang et al., 2017), MobileNetV2 (Sandler et al., 2018), ResNet18 (Wu et al., 2018), ResNet50 (He et al., 2016) , ResNet101 (He et al., 2016), Xception (Chollet, 2016), InceptionResNetV2 (Ioffe and Szegedy, 2015; Szegedy, Ioffe, et al., 2016), ShuffleNet (Zhang et al., 2017), and NasNetMobile (Zoph et al., 2018). In the third step, traditional feature extractors, PSDNet, and pretrained ConvNet were evaluated on the real photograph dataset in two ways. First, the dataset was split into the regular training/validation/test subdatasets. Second, one material was removed completely during training and validation to check the ability of ConvNet to predict the PSD of materials that the network has never seen. The unseen material was used as the test dataset.
Our findings indicate that each of the compared techniques can lead to good predictions of the PSD for both real and synthetic datasets. For the synthetic dataset, the Root Mean Square Error (RMSE) value on the percentages passing was 3.4 % when using a selection of the 618 best traditional features as the input for an artificial neural network. The best version of PSDNet resulted in a RMSE value on the percentages passing of 2.8%. InceptionResNetV2 and DenseNet201 gave the best results for feature extraction and transfer learning, respectively, with the same RMSE values of 3.6 %. For the real photographs, PSDNet achieved a RMSE on the percentages passing of 3.7 and 5.5 % for the color and grayscale band, respectively. The best performances for pretrained models' transfer learning and feature extraction were obtained for NasNetLarge and ResNet101 with a RMSE of 4.3 and 4.0 %, respectively. The best results were obtained by using InceptionResNetV2 as the transfer learning-based feature extraction method, with an RMSE value of 1.7 %. As well, similar results were obtained by using all PSDNet gray and color extracted features consisting of 1000 components, with an RMSE of 1.8%.
Significantly better results were achieved when using color images in PSDNet instead of grayscale images. Better results were also obtained for the fine particles in the PSD, especially for traditional feature extraction techniques. A combination of views of the soil captured from the top and bottom yields better results for PSD prediction than using only the top view. For the real photographs, good results were achieved in the material removal test with a minimum RMSE of 1.5 %. Much higher RMSE values (> 31.8 %) were obtained for the two extreme PSD (coarsest and finest). This highlights the fact that the methods that were compared in this thesis cannot be used to extrapolate outside of the PSD range used in the training dataset.
This study was the first to introduce pre-trained ConvNet for feature extraction or transfer learning methods of PSD prediction. The training with a real dataset is a key difference between PSD determination with neural networks and classical image processing methods. Training allows hidden particles to be taken into account implicitly. Commercial software based on particle segmentation must account for these hidden particles explicitly using statistical methods. The capacity of ConvNet to learn also allows the model to be enhanced during its operation through the extension of the dataset. The synthetic dataset that was used in this paper is seen as a promising way to build large datasets including both real photograph and synthetic images.
| Date | 24 May 2022 |
|---|
| Original language | American English |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | François Duhaime (Supervisor) |
|---|