A robust defect detection method with a generalization enhancer and cross-modality aggregator for cylinder bores

Abstract

High-quality cylinder bores in automobile engines enable drivers to respond quickly to emergencies. Automated detection methods are gradually being adopted across various industries. However, uncontrollable factors and improper preservation methods lead to various types of defects on cylinder bores, thereby causing existing high-performance detectors to exhibit not only undesirable generalizability for unseen defect types but also a certain degree of missed detection for defects of seen types, thereby allowing defective cylinder bores to flow into the market. To address these issues, we propose a foundation-model-based robust defect detection method with high generalizability for cylinder bores (RHG-Detector). Specifically, to address unseen defect categories, we propose a generalization enhancer comprising a box filter, a region extractor and a defect discriminator (DeDi) based on a foundation model to extend defect detection from a closed set to an open set. To reduce missed detections, we adopt a cross-modality aggregator to aggregate the detection results from different modalities. Additionally, we collected and annotated challenging defect classification and detection datasets for cylinder bores, named HIT-EngDC (Harbin Institute of Technology Engine defect classification dataset) and HIT-EngDD2 (Engine defect detection dataset-version 2), which cover nearly all types of cylinder bore defects. Extensive experiments on HIT-EngDC and HIT-EngDD2 demonstrate the state-of-the-art performance of RHG-Detector, with a classification accuracy of 92.0 and a mAP@50 (mean average precision under intersection over union = 0.5) score of 45.2, where the latter is increased by ∼6 and ∼3 compared to the corresponding FasterRCNN (faster region-based convolutional neural networks) and YOLOv7 (you only look once) scores, respectively.