Crack detection and recognition model of parts based on machine vision

The traditional manual crack detection method for parts is inefficient, subjective, and has low accuracy. Laser and radar equipment is accurate but costly. Thus, image processing methods based on machine vision are widely used owing to the rapid detection speed and low cost. However, these methods are inappropriate to a low-contrast, high-noise, and poorresolution environment. To obtain accurate detection results in complex environments, this study proposed a new automatic crack identification model. Image quality was enhanced using the multi-scale Retinex algorithm modified by a three-sided filter. The algorithm for spectral clustering was employed to segment the cracks, thereby potentially preserving the relevant information between the samples. Pixel interference regions were removed on the bases of shape feature and contour information. Target feature vectors were extracted according to the characteristics of the crack, and five features of the crack image were constructed. The SVDD(support vector data description) classifier was constructed by using machine learning methods to automatically distinguish between crack and non-crack structures. The crack direction was identified on the basis of the projections in the vertical and horizontal directions. The accuracy of the model was verified by experiments. Results indicate that the crack images with a size of 300 x 400 have good segmentation effect and low complexity through the clustering algorithm, while the cluster number is equal to 2. Noise interference can be further reduced by setting the crack area between 2% to 28%. Moreover, the detected crack width can be accurate to 1 pixel. The crack and non-crack areas can be classified using the SVDD classifier without large samples. The proposed method provides good reference for inspection platform construction and state assessment. © 2019 School of Science, IHU. All rights reserved.