Vision-Based Instance Segmentation of Piezoelectric Crystal Defects Using a Hybrid Bidirectional CNN-Transformer...

Accurate defect instance segmentation is essential for automated quality inspection of piezoelectric crystals, where scratches, stains, and other surface defects often show large scale variation, irregular morphology, and weak contrast against complex backgrounds. Existing CNN-based segmentation models capture local texture effectively but are limited in global context modeling, whereas Transformer-based models often require stronger local detail preservation for fine industrial defects. To address these limitations, this paper proposes a hybrid bidirectional bridging CNN-Transformer network, named HBCTNet, for piezoelectric crystal defect instance segmentation. HBCTNet introduces a Convolution-Transformer Bidirectional Bridging Block (CTB) that enables lightweight two-way interaction between convolutional local features and Transformer global representations. A Frequency and Spatial Convolution (FSC) module is designed to enhance defect-related details in both spatial and frequency domains, and a Multi-Scale Feature Pyramid Network (MS-FPN) fuses features across scales. On the reported image partition of the self-constructed dataset, HBCTNet-S obtains 92.3% box AP and 89.9% mask AP with 12.67 M parameters and 48.1 G FLOPs. Its mask AP is 2.1 percentage points higher than YOLO11-SEG and 2.4 percentage points higher than YOLOv9-SEG under the reported comparison. These results support the effectiveness of the architecture within the evaluated dataset; inference speed and cross-domain generalization remain to be assessed.