ya zhao ^* yu guan di jia ^*

• Northeast Petroleum University, Daqing, China

Residual oil segmentation plays a crucial role in assisting researchers in predicting hydrocarbon distribution and analyzing residual oil properties in pore space. In recent years, the introduction of deep learning technology has greatly promoted the progress of microscopic residual oil recognition technology. However, during the acquisition of microscopic residual oil images, the boundaries between residual oil, water, and rock matrix in the images are often unclear due to objective factors such as the field environment, the quality of core preparation, and the operation method, which significantly increases the difficulty of segmentation. Existing recognition methods are often difficult to achieve accurate segmentation when dealing with tiny residual oil targets with fuzzy edges and discrete distribution.To address the above challenges, this study improves the DeepLabV3+ model to achieve accurate segmentation of microscopic residual oil images. We chose the lightweight MobileNetV2 as the backbone network to balance the segmentation accuracy and processing speed; meanwhile, we introduced a multi-scale strip pooling module in the ASPP module to accurately capture the location information and remote dependencies of the remaining oil. Furthermore, a genetic algorithm has been utilized for automatic optimization to identify the optimal dilation rate configuration. Finally, by building a feature-guided fusion module, we achieve lightweight and efficient fusion of different hierarchical and semantic features.Experimental results on the homemade micro residual oil dataset show that our improved model achieves an accuracy of 95.59% and an MIoU of 91.10%, both of which exceed the current mainstream segmentation models. In addition, we validate the improved method on the public dataset BUSI, which also demonstrates high accuracy.Moreover, the proposed algorithm has only 4.65M parameters and takes only 7.5ms to segment an image, which combines high segmentation accuracy and operation efficiency, demonstrating its strong potential in the microscopic residual oil recognition task.