A Measurement System for the Tightness of Sealed Vessels Based on Machine Vision Using Deep Learning Algorithm

Tightness defects on sealed vessels, such as filters, may cause serious environment pollution and potential safety hazards, which means that the tightness measurement of sealed vessels cannot be neglected. For the measurement of microleakage, the traditional methods are greatly affected by the ambient temperature, leading to unstable results. In this article, a novel mechanism and method based on deep learning for tightness detection and quantification of the sealed vessels is proposed. First, you only look once (YOLO)v5 network with asymmetric convolution blocks (Ac.Bs) in the backbone network is applied to tightness measurement, which improves the feature extraction capability of small targets. Second, a filling algorithm for eliminating crack (FEC) is reported. In this algorithm, novel horizontal and vertical marking operators are defined, which can accurately obtain geometric and motion parameters of the bubble. Third, a calculation model is established to calculate the volume of the bubble quickly under the premise of known bubble area and motion parameters. Fourth, an automatic dry-type measuring device for measuring leakage has been developed to provide an experimental platform for the measurement framework. Finally, performance testing is performed on an independent dataset. The mean intersection over union (mIoU) of the proposed bubble detection method is 98.74%, the processing time for a single image is 6 ms, and the measurement precision of the system is 0.03 mL. The experimental results demonstrate that the proposed tightness detection mechanism and method can greatly improve the accuracy and stability of tightness detection of sealed vessels, which have good comprehensive performance.