Abstract:

The intelligent identification of shale scanning electron microscope (SEM) images can rapidly analyze shale reservoir minerals, which is one of the important means of predicting the “sweet spot” of shale oil reservoirs, and is also a future technological development trend. Traditional methods have problems such as low automation, low sample suitability, and limited feature extraction when identifying mineral components. To this end, this paper proposes a BlendMask-based SEM image characterization method for shale. Firstly, image preprocessing techniques such as bilateral filtering, Laplacian, and image normalization are used to denoise, sharpen, and unify the pixel of original images to improve the quality of training samples; Then, image augmentation methods such as rotation, scaling, and luminosity change are used to construct augmentation strategies to expand the number of datasets; And finally, the BlendMask network is improved by using the attention mechanism and the depth separable convolution which is used to realize the component segmentation and recognition of images. The experimental results of shale SEM images applied to Haita basin show that the segmentation accuracy and recall of the improved method are improved by 0.02-0.20 and 0-0.59, respectively, and the segmentation time is reduced by 1.29-2.70 s compared to the BlendMask model.

Key words: shale oil “sweet spot” reservoirs')">

shale oil “sweet spot” reservoirs, BlendMask')"> BlendMask, scanning electron microscope images, mineralogical composition, segmentation and identification