Gravel Extraction from FMI Based on DSAM-DeepLabV3+ Network

Accurate representation of the gravels plays a fundamental role for fine characterization and comprehensive evaluation of glutenite reservoir. Formation microscanner images (FMI) is an advanced well log method, which can provide intuitionistic 2D images of the rocks underground. The existing applications of FMI for reservoir evaluation relies on manual fitting or subjective judgment by experts, which is inefficient and lack of accuracy. Several attempts so far have been made to extract useful information from FMI such as grayscale threshold segmentation, connected domain technology and wavelet transform image segmentation technology. However, the above methods are either sensitive to noise or easy to cause over-segmentation, so the segmentation results are always unsatisfactory. Recently, in image processing area, DeepLabV3+ network has become a powerful tool for image semantic segmentation by combining spatial pyramid pooling with dilated convolutions to capture the high-level semantic information. To this end, in this paper, we propose to apply DeepLabV3+ network for gravel extraction by segmenting the FMI. In order to achieve better pixel-level recognition, we design a dual serial attention module (DSAM) and embed it into the conventional DeepLabV3+network, which can capture a large number of low-level image statistical information. The improved DeepLabV3+ network is called DSAM-DeepLabV3+. Experimental results show that DSAM-DeepLabV3+ pushes the performance of gravel extraction further.