Research on Detection Method of Coating Defects Based on Machine Vision

Aiming at the problems in the current coating surface defects detection that it is difficult to characterize the defect features, furthermore the detection accuracy and efficiency are hard to meet industrial demand, in this paper, a machine vision system for coating defects detection is designed; then, a coating defects classification method based on convolutional neural network which is trained and tested through cross-validation to realize the classification of multi-type coating defects, is proposed. According to the collected coating dataset including defect-free coating and four types coating defects: crack coating, running coating, orange peeling coating and adhesion failure coating, the classification performance of multi-type convolutional neural networks is analyzed experimentally. Among the five convolutional neural networks, Resnet50 achieves the best detection effect, precision: 95.0% and accuracy: 97.9%. The detection performance of Densenet121 is similar to Resnet50’s, but the model size of Densenet121 is only 1/3 of former’s; furthermore, these two types of networks are tested on captured coating defects in actual spraying process, the average precision and accuracy of classification were 93.3% and 97.3%, 91.8% and 96.7%, respectively, and the detection time for each image is 0.028s and 0.025s, respectively. Therefore, Experiments prove that the purposed method is convenient and quick to detect coating surface defects, and it has high precision and accuracy. Thus, the method can be used for industrial site detection.