Borehole image inpainting based on multidimensional and multi-scale dilated transformer

Abstract From borehole image, complex formation conditions of the borehole wall can be obtained with high-resolution and important parameters such as holes, fractures, and layers can be extracted for subsequent reservoir evaluation. However, due to the structural characteristics of wellbore imaging logging tools and the limitations of logging methods, some abnormal phenomena such as blank stripes, misaligned electrode plates, and black-and-white scratches appear on borehole image and severely affect the image quality required for subsequent logging evaluation. Therefore, in this study, a multidimensional, multi-scale dilated transformer model (MMDT) was proposed for large-area masked image inpainting. Based on the multi-head self-attention mechanism, the model was adaptively designed for the grayscale characteristics and the fixed size of borehole image. Then, feature concatenation instead of residual learning was employed to mitigate the gradient explosion problem of the original model. Additionally, style module was introduced to make the texture of inpainted regions more natural. Finally, multi-scale dilated convolution was incorporated to further enhance semantic associations. Experimental results showed that MMDT improved the Fréchet inception distance evaluation metric has improved by 56.86%, 46.29%, 29.74%, and 17.88% compared to four algorithms (LaMa, DeepFillV2, FILTERSIM, and Criminisi), respectively. The inpainting effect of abnormal phenomena was also significantly improved by MMDT compared to other algorithms.