Machine Learning for Cyber Security: Third International Conference, ML4CS 2020, Guangzhou, China, October 8–10, 2020, Proceedings, Part III

Scientifically, the identification of salt ore has definite practical significance for the exploitation of oil and gas. Traditionally, this is achieved by picking the salt boundaries with manual vision, which may introduce serious systematic bias. Nowadays, with the technological progress of machine vision used in image analysis, human effort has been replaced by machine capacity in salt mine recognition. Especially, with the in-depth application of deep learning technology in the field of machine vision, salt mine recognition using image analysis is revolutionizing with more acceptable efficiency and accuracy. To this end, with exploratory data analysis to mine the characteristics and data processing to increase the size of the image data for further enhancing the generalization capability of the designed model, a deep convolutional neural network based image segmentation model is investigated to achieve salt mine recognition in this paper. Concretely, a UNet model integrated modified ResNet34 is first designed as a basic recognition model, and many attempts then are conducted to further optimizing the model according to the data characteristics, including adding auxiliary function, hypercolumn, scSE and depth supervision scheme. In addition, multiple loss functions are also attempted to be adapted to further improving the model generalization capacity. The numerical analysis and evaluation finally show the efficiency of the investigations on loss value and recognition accuracy.