Abstract
Fullbore Formation MicroImager (FMI) logging provides high-resolution electrical images of the borehole wall and has been widely used for fracture interpretation, dip analysis, and complex reservoir evaluation. However, due to limited pad coverage, borehole diameter variations, and unstable tool–wall contact, original FMI images commonly contain longitudinal blank stripes along the depth direction. These missing regions disrupt the continuity of key geological features, such as bedding interfaces, fractures, and vugs, thereby reducing the reliability of subsequent geological interpretation. Conventional methods, including inverse distance weighting and intensity-space interpolation, are computationally efficient but mainly rely on local smoothing assumptions, making them insufficient for restoring structural trends across wide missing stripes. Multiple-point geostatistical methods depend strongly on training images and may produce overly rigid texture patterns in complex heterogeneous intervals. Deep learning-based inpainting methods, particularly those based on generative adversarial networks, usually require large-scale training datasets and may introduce geologically inconsistent hallucinated details. To address these limitations, this study proposes a zero-shot FMI blank-stripe filling method based on Deep Image Prior (DIP) and an attention-enhanced U-Net embedded with Convolutional Block Attention Modules (CBAM). The proposed method requires no external training data. Instead, it uses fixed random noise as input and optimizes the CBAM-enhanced U-Net generator through masked self-supervision over valid pixels. In this way, the structural prior of the network is exploited to extend geological textures from the observed regions of the current FMI image into the blank stripes. To further alleviate brightness drift and boundary artifacts during zero-shot optimization, depth-adaptive intensity correction and Poisson blending are incorporated. In addition, a chunk-wise scheduling strategy is adopted to ensure stable restoration of long logging image sequences. Experimental results demonstrate that the proposed method improves the geometric continuity of sinusoidal bedding traces and fracture textures within blank stripes while preserving the physical intensity consistency of FMI images. Compared with conventional interpolation, multiple-point geostatistical methods, and GAN-based inpainting approaches, the proposed method is more suitable for FMI image restoration tasks that require high geological reliability.
Keywords: FMI logging; Blank-stripe filling; Deep image prior; U-Net; Poisson blending; Zero-shot learning
