关键词: 矿石图像分割, 语义分割, 线性可变形卷积, 注意力机制, 边界检测, 双域协同

Abstract: In order to solve the problems of boundary blurring and insufficient accuracy caused by complex texture, irregular shape and uneven illumination of ore images in the segmentation task, a segmentation network with linear deformable convolution and dual-domain synergistic dynamic attention—LDDA-Net was proposed. The network adopts the encoder and decoder architecture, and constructs an adaptive sampling point distribution through Linear Deformable Convolution (LDConv) in the serial dual-feature encoder, which flexibly fits the irregular shape of the ore, and effectively controls the computational overhead with its linear characteristics. Secondly, a dynamic attention modulation module (DAM Module) was designed for spatial domain features, which realized the dynamic focusing and reinforcement of key information in the feature map and the edge of the ore through pooled sampling, learnable attention matrix and boundary-sensitive weight allocation mechanism. Finally, a new dynamic progressive attention-guided loss function (DPAG Loss) is proposed, which guides the model to focus on hard-to-divide areas such as fuzzy boundaries and small particles of ore during the training process by dynamically generating attention maps in multiple stages, and forms a space-loss dual-domain synergy with the DAM module, and constructs a feedback closed-loop mechanism of feature perception and learning strategies. The experimental results on the self-built OpenPitOre dataset and the public ore dataset show that the HD95 boundary error of LDDA-Net is only 16.84mm, which is 11.37% lower than that of the suboptimal model VM-Unet. The Dice coefficient is as high as 91.54%, and the mIoU and PA are 85.13% and 94.10%, respectively, which are significantly better than the comparative segmentation models, which fully verifies the high-precision and fine segmentation effect achieved by LDDA-Net in complex scenes, and provides reliable technical support for the intelligent detection and blockiness analysis of open-pit blasting ore.

Key words: ore image segmentation, semantic segmentation, linear deformable convolution, attention mechanisms, boundary detection, Dual-domain synergy