A weakly supervised cascade segmentation method for thin core interbeds using dual-channel attention and adaptive PSO

Abstract

Precise localization and quantitative characterization of thin interbedded targets in core images are essential for revealing reservoir heterogeneity and predicting remaining oil distribution. Existing fully supervised deep learning methods rely heavily on costly pixel-level annotations, while repeated downsampling often leads to the loss of high-frequency details in extremely thin layers. To overcome these inherent limitations, this paper introduces a weakly supervised ‘classification–localization–segmentation’ cascade framework, achieving fine-grained, pixel-level representations relying solely on image-level annotations. First, a channel–spatial multi-scale fusion network (CSMFN) incorporating a dual-pooling coordinate attention mechanism is developed for classification. To prevent small-target features from being overwhelmed by background information in conventional global average pooling, a max-pooling branch is introduced to capture high-frequency texture extremes. Jointly with average pooling, the design suppresses background noise and enhances microbedding feature responses. Subsequently, a hybrid paradigm of ‘semantic coarse localization–physical fine segmentation’ is established. Class activation mapping generates high-confidence masks to restrict the regions of interest, while particle swarm optimization (PSO) is deployed to dynamically compute the most suitable segmentation thresholds within the noise-reduced grayscale domain. By leveraging grayscale mutation characteristics of geological targets, the method effectively avoids boundary smoothing caused by continuous downsampling in end-to-end networks. Experiments on the Hade Oilfield dataset achieved an F1 score of 93.5% with an inference time of 72 ms per image. Despite lacking pixel-level annotations, the proposed method outperformed the fully supervised DeepLabV3 + in edge localization of ultra-thin interlayers. Quantitative analysis further characterized heterogeneity in the Donghe Sandstone Member, providing reliable support for geological evaluation.