Improving pipe perforation estimates from ultrasonic imaging using subpixel machine learning trained on optical data

During well completion, well pipes are perforated to gain access the reservoir. The size and shape of the perforated holes can be found from these holes’ outlines, which are generally estimated by optical or ultrasonic image logging. While optical imaging typically has a much higher spatial resolution than ultrasonic imaging and thus allows more precise outline estimates, optical imaging requires transparent liquids inside the pipe. Ultrasonic imaging, on the other hand, can be performed in a wider variety of liquids and can provide further information about the well state from the same measurement. One strategy is therefore to combine both types of measurement on the same toolstring. Thus, ultrasonic imaging can take over the job of estimating hole outlines from optical imaging when the liquid is no longer sufficiently transparent. One issue with this strategy is that the agreement between these two imaging techniques currently leaves much to be desired. This work addresses this issue by training a machine learning (ML)-based subpixel segmentation algorithm to take ultrasonic images of perforations and reproduce perforation outline estimates made from optical images. This approach assists the algorithm in drawing out information from the ultrasonic data which is not easily accessible using traditional image processing techniques. We use a dataset of 390 perforations, measured by both an optical and an ultrasonic tool, to train and test the machine learning algorithm. For comparison, we use a baseline algorithm based on interpolation and image thresholding. We evaluate the algorithms’ performance according to their estimated outlines’ match with the optical outlines. The outlines’ overlap is quantified via the intersection over union metric (baseline: 50.8 %, ML: 74.2 %; higher is better), and their area match is quantified via mean relative area error and compared to the results of another study from the literature (baseline: 54.4 %, ML: 18.7 %, other study: 54.5 %; lower is better). • We present a machine learning algorithm to segment pipe perforation holes. • The algorithm reproduces optical segmentations from ultrasonic data. • It vastly outperforms a baseline algorithm based on traditional image processing. • We find better performance than seen in previously reported work in the literature.