Froth texture extraction with deep learning

[摘要] ENGLISH SUMMARY: Soft-sensors are of interest in mineral processing and can replace slower or more expensive sensors by using existing process sensors. Sensing process information from images has been demonstrated successfully, but performance is dependent on feature extractors used. Textural features which utilise spatial relationships within images are preferred due to greater resilience to changing imaging and process conditions.Traditional texture feature extractors require iterative design and are sensitive to changes in imaging conditions. They may have many hyperparameters, leading to slow optimisation. Robust and accurate sensing is a key requirement for mineral processing, making current methods of limited potential under realistic industrial conditions.A platinum froth flotation case study was used to compare traditional texture feature extractors with a proposed deep learning feature extractor: convolutional neural networks (CNNs). Deep learning applies artificial neural networks with many hidden layers and specialised architectures for powerful correlative performance through automated training. All information of the input data structure is determined inherently in training with only a limited number of hyperparameters. However, deep learning methods risk overfitting with small datasets, which must be mitigated.A CNN classifier and a framework for unbiased comparison between feature extractors were developed for predicting high to low grade classes of platinum in flotation froth images. CNNs can perform all the functions of a soft-sensor, but this may bias performance comparison. Instead, features were extracted from hidden layers in CNNs and fed into a traditional soft-sensor. This ensured performance measurements were unbiased across all feature extractors. With a full factorial experiment, the following CNN hyperparameters were evaluated: batch size, number of convolutional filters, and convolutional filter size.Accuracy of grade classification was used to score feature extractors. These reference texture feature extractors were compared to CNNs: Local Binary Patterns, Grey-Level Co-occurrence Matrices, and Wavelets. The impact of spectral features (bulk image features such as average colour) was also evaluated, as CNNs can also use spectral image properties to create features, unlike traditional texture extractors. Extractors were tested with input resolutions from 16x16 to 128x128 with two soft-sensor models: Linear Discriminant Analysis, and k-Nearest Neighbour classifiers. Optimal grade classification accuracies were: CNN – 96.5%, LBP – 100%, GLCM – 73.7%, Wavelets – 98.3%, and Spectral – 98.4%Training CNNs to extract features was successful with robust results regardless of hyperparameters selected. The only statistically significant differences obtained during training were that smaller batch size and smaller input resolution gave superior training performance. Results were found to be reproducible for all models. Analysing learned CNN features indicated both textural and spectral features were utilised. Overall results showed spectral features gave good classification performance, potentially adding to CNN performance. CNNs showed comparable performance to other texture feature extractors at all resolutions.This proof of concept implementation shows promise for deep learning methods in mineral processing applications. The resilience of CNNs to changes in imaging and process conditions could not be evaluated due to limited data in the case study. Future work with deep learning methods, while promising, will require larger datasets which are more representative of a variety of process conditions.

[发布机构] Stellenbosch University