Automatic segmentation of breakouts in acoustic borehole image logs using convolutional neural networks

Resumo

Breakouts are collapsed zones on borehole walls caused by compressive failure. The identification
of breakouts in wellbores is fundamental for estimating the stability of the well and to obtain the magnitude of
the maximum horizontal stress present in the rock formation. Traditionally, professional interpreters identify and
characterize breakouts manually, which can be considered a very time-consuming task due to the massive size of
the borehole data. Due to the complexity of the structures of interest and several noisy artifacts in the image log,
traditional image processing methods are not very effective in solving the problem. The U-Net proposed by Olaf
Ronneberger et al. is a convolutional neural network model commonly used in medical image segmentation that
has been applied on several areas. This architecture is composed of two parts: the encoder, which is used to capture
the image context, and the decoder, which is used to allow a precise location using transposed convolutions. A
series of changes in this architecture has been proposed to improve the network’s capacity to extract features in
multiple scales and improve the skip connections between encoder and decoder. The DC-UNet (Dual Channel
U-Net) is one of the U-Net based models designed to overcome some limitations of the original network. In
this work, we present the application of DC-UNet for breakout segmentation in wellbore’s amplitude image logs.
Furthermore, we also discuss the problems related to the inaccuracy of the data’s annotated masks and the image
pre-processing strategies applied to reduce their effects.