Geometric spatial constraints network for slender and tiny surface defect detection

Abstract

Detecting defects on aircraft impeller surfaces is challenging due to the thin and fragile structure of certain defects, as well as their varying scale and geometry. To address these two challenges, we propose the Geometric Spatial Constraints Network (GSCNet) for precise impeller defect detection. First, we develop an automatic image acquisition equipment to capture high-quality data of impeller surface defects. Subsequently, we introduce GSCNet, which comprises two main components: Rich Semantic Information Representation (RSIR) and Spatial Correlation Awareness (SCA) to detect surface defects. Within RSIR, we propose a geometric-constraints-guided, deformable-convolution-based module named Slender Partial Convolution (SPC), along with a Multi-Geometric Feature Fusion (MFF) module. SPC captures the features of tubular structures without redundant information by aligning the convolution kernel shape with slender defects, while MFF facilitates the fusion of various semantic features, thereby enhancing the ability to extract semantic information. In SCA, we introduce a novel attention mechanism that captures inherent spatial correlation to enhance the high-similarity defects classification capability by modeling representative spatial information. Finally, we design a similarity-enhanced loss function to further improve the detection of multiple geometric defects simultaneously, as it alleviates the scale sensitivity of IoU-based loss. Comparative experiments demonstrate that our framework outperforms all representative detection models, achieving 83.2% mAP on the AISD dataset, which surpasses the second-best model by 3.8%. The first set of ablation experiments confirms the effectiveness of each module within the framework. The second set of ablation experiments on the NEU-SEG and MT datasets validates the feature extraction and plug-and-play capability of RSIR. The generalization ability of GSCNet is further demonstrated on the NEU-DET and GC10-DET datasets.