Intelligent identification of fractures and holes in ultrasonic logging images based on the improved YOLOv8 model

Aiming to address the demand for intelligent recognition of geological features in whole-wellbore ultrasonic images, this paper integrates the YOLOv8 model with the Convolution Block Attention Module (CBAM). It proposes an intelligent method for detecting fractures and holes, as well as segmenting whole-wellbore images. Firstly, we develop a dataset sample of effective reservoir sections by integrating logging data and conducting data augmentation on fracture and hole samples in ultrasonic logging images. A standardized process procedure for the generation of new samples and model training has been proposed effectively. Subsequently, the improved YOLOv8 model undergoes a process of training and validation. The results indicate that the model achieves average accuracies of 0.910 and 0.884 in target detection and image segmentation tasks, respectively. These findings demonstrate a notable performance improvement compared to the original model. Furthermore, a sliding window strategy is proposed to tackle the challenges of high computational demands and insufficient accuracy in the intelligent processing of full-well ultrasonic images. To manage overlapping regions within the sliding window, we employ the Non-Maximum Suppression (NMS) principle for effective processing. Finally, the model has been tested on actual logging images and demonstrates an enhanced capability to identify irregular fractures and holes, which significantly improves the efficiency of geological feature recognition in the whole-well section ultrasonic logging images. • This paper proposes a deep learning-driven method for detecting geological features of full-wellbore ultrasonic images. • We develop a sample dataset of geological features and conduct data augmentation on fracture and hole samples. • We have integrated the Convolution Block Attention Module (CBAM) into the YOLOv8 model. • A sliding window strategy to achieve intelligent detection throughout entire wellbore sections effectively. • A standardized process for the model training of additional geological features under different well conditions.