IDDM: An incremental dual-network detection model for in-situ inspection of large-scale complex product

Abstract

Appearance inspection is crucial for quality control during the manufacturing of large complex products. In-situ visual inspection based on image processing and machine learning can significantly reduce the production costs by avoiding the stages of transshipment and relocation, etc. However, imaging and model training are challenged by the complex background, illuminance, and changing environment of the production site. In addition, large dataset is hard to be obtained with object-level annotations owing to the high cost of manual annotation in practical situations. In this paper, we proposed an Incremental Dual network Detection Model (IDDM) for efficient and high-precision inspection of the appearance of large complex product base on in-situ images. A dual network structure is used to implement the incremental training of the model based on metric learning and batch labeling of unlabeled data during the training on a small number of well-labeled samples. The regions of interest are extracted based on multi-feature and refined to improve the positional accuracy of the defects in complex background. On the public dataset, the experimental results derived on various annotation scales showed a better performance of the proposed IDDM compared to the supervised object-detection baselines. The mean Average Precision (mAP) was 51.8% with a 25% labeled ratio and the processing speed was 22 frames per second (FPS). In addition, the proposed method was applied to the defect detection of the snow groomer surface as an industrial case.