Abstract: In the field of oil development, accurate identification and classification of microscopic remaining oil is crucial for improving the oilfield exploitation efficiency and oil recovery. However, the traditional remaining oil identification technique is confronted with the problems such as low identification efficiency, low accuracy and high resource consumption, thus restricting its practical effectiveness in oilfield applications. Therefore, LLGFormer, a microscopic remaining oil image classification network based on convolutional neural network (CNN)and vision transformer (ViT), is proposed, which can not only significantly enhance the classification accuracy but also improve the operational efficiency by fusing local and global features. Firstly, an edge perception enhancement module is designed to enhance image edge texture information, and then the local and global features of remaining oil are extracted in parallel by LLGFormer block. In addition, a contribution discriminant network is introduced to guide the ViT branch to focus on effective information, and a step-by-step computation strategy is adopted to reduce the calculation amount of the model. Moreover, the validity of LLGFormer is verified by experiments on the homemade microscopic remaining oil dataset and the public dataset. This not only proves the significant advantages of LLGFormer in terms of the balance between the speed and performance of microscopic remaining oil image processing, but also provides a new technical path for the automated identification and classification of microscopic remaining oil in the petroleum industry.