Small Sample Multivariate Image Regression with Convolutional Neural Networks

Multivariate image regression has important applications in mining, mineral processing and manufacturing, where the collection of image data is becoming an integral part of process operations. Deep learning models, and convolutional neural networks in particular, can provide better models than traditional approaches, but their rapid adoption has also brought new challenges, of which the scarcity of reliable labeled data is a major bottleneck. In this study, multivariate image regression on a coal image data set with seven target values only and assays from arsenic froth image data with 30 target values were considered. It is shown that pretrained deep convolutional neural networks can directly generate features that are competitive with traditional methods. On both data sets, they could directly extract features from the images that yielded models as or more reliable than what could be obtained with engineered features sets based on grey level co-occurrence matrices, latent variables and local binary patterns. Fine-tuning yielded further significant improvement on both data sets.