A rock-type object detection method for petroleum exploration incorporating a hybrid aggregation network

Abstract

Rock-type identification is essential in petroleum exploration and geological engineering, directly supporting lithology analysis, stratigraphic correlation, and reservoir evaluation. Traditional expert-driven approaches are inefficient and subjective, limiting their applicability to large-scale automated analysis. Although deep learning has been applied to rock image recognition, most studies rely on image-level classification and fail to exploit the engineering advantages of object detection. DEIM, a Transformer-based end-to-end detector, improves training efficiency through dense one-to-one matching and a matchability-aware loss, but its convolutional backbone limits feature representation for complex textures and fine-grained rock categories. To overcome these limitations, this paper proposes OE-DEIM, which introduces a Mixed Aggregation Network (MANet) before the Transformer encoder without modifying DEIM’s original matching mechanism or loss function. MANet enhances multi-scale feature representation and channel–spatial coupling by combining standard convolution, depthwise separable convolution, and multi-branch aggregation. Experiments on a self-constructed dataset with nine rock categories show that OE-DEIM outperforms the original DEIM in both detection accuracy and training stability, demonstrating its effectiveness for intelligent rock-type recognition.