• We introduce DMRF-Net, a robust framework designed for scale-variant target detection in industrial settings. It utilizes deformable large-kernel decomposition attention to model target position offsets while integrating global and local context semantics. • We introduce the Context-Aware Deformable Kernel Module (C3K2), which uses large-kernel decomposition attention to precisely model target position offsets while effectively merging global and local context information. • We introduce the Progressive Semantic Aggregation (PSA) module, which replaces traditional static attention with a dynamic multi-layer routing mechanism. This approach refines fine-grained semantic associations, enabling better detection of minor defects. • We introduce a large-kernel feature aggregation module that integrates dual-path adaptive pooling to combine multi-scale local features with global semantics. This module constructs a spatially aware feature correlation network, improving target detection across varying scales. In the monitoring of polysilicon reduction furnaces, YOLOv-based methods have garnered significant attention due to their ability to strike a balance between efficiency and accuracy, making them particularly suitable for industrial automation scenarios that require real-time processing and large-scale edge deployment. However, previous YOLOv-based target detection algorithms often face challenges, such as ignoring the edge features of subtle defects and insufficient adaptability to dynamic environmental changes. These limitations result in a poor generalization ability of the detection model and insufficient detection accuracy for multi-scale targets. To address these issues, this paper proposes a multi-scale target detection network to improve the model’s ability to locate and represent targets semantically, namely DMRF-Net. First, a depthwise deformable large kernel decomposition attention mechanism is embedded into the context-aware layer to perceive the offset of the target position. Simultaneously, an interaction is established between global and contextual semantics to reduce the representation of redundant information. Second, a progressive semantic aggregation module with a novel multi-layer routing attention embedding is employed, replacing the traditional static attention mechanism with a dynamic hierarchical filtering mechanism. This approach constructs a progressive semantic association from coarse-grained to fine-grained, enabling the detection of subtle defects. Furthermore, this paper develops a large-kernel spatial awareness selection and aggregation module using dual-path adaptive pooling, which combines multi-scale local features with global semantics to detect subtle changes in target scales in industrial environments sensitively. Experimental results on real industrial datasets demonstrate that the proposed DMRF-Net significantly improves performance in high-precision scenarios (e.g., polysilicon reduction furnace monitoring and steel plate surface defect detection), achieving mAP@0.5 of 0.86 and 0.82, respectively.
Sun et al. (Sun,) studied this question.