Optimization of small-sample microscopic residual oil distribution recognition based on Mask R-CNN and migration learning

Artificial intelligence has undergone swift advancements in recent years, and within the petroleum industry, the identification of microscopic residual oil distribution is crucial to the accurate extraction of hydrocarbons and the enhancement of recovery rate. However, the traditional recognition methods have insufficient accuracy in small sample scenarios, and it is difficult to effectively recognize the residual oil distribution in complex regions such as low-permeability layers and the upper part of thick oil layers. Therefore, the research proposes a small-sample microscopic residual oil distribution identification model based on masked region Convolutional Neural Network and transfer learning. Experimental results indicated that the microscopic residual oil distribution recognition model achieved a recognition accuracy of 72.64%, with a precision of 71.92% and an F1 score of 72.05%, outperforming the comparison model by a considerable margin. The model can realize high-precision automated recognition under small samples and provide scientific data support for oilfield residual oil dredging and development plan optimization.