Find in Library

The quality, wear and safety of metal structures can be controlled effectively, provided that surface defects, which occur on metal structures, are detected at the right time. Over the past 10 years, researchers have proposed a number of neural network architectures that have shown high efficiency in various areas, including image classification, segmentation and recognition. However, choosing the best architecture for this particular task is often problematic. In order to compare various techniques for detecting defects such as “scratch abrasion”, we created and investigated U-Net-like architectures with encoders such as ResNet, SEResNet, SEResNeXt, DenseNet, InceptionV3, Inception-ResNetV2, MobileNet and EfficientNet. The relationship between training validation metrics and final segmentation test metrics was investigated. The correlation between the loss function, the <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>D</mi><mi>S</mi><mi>C</mi></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>I</mi><mi>o</mi><mi>U</mi></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>R</mi><mi>e</mi><mi>c</mi><mi>a</mi><mi>l</mi><mi>l</mi></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>P</mi><mi>r</mi><mi>e</mi><mi>c</mi><mi>i</mi><mi>s</mi><mi>i</mi><mi>o</mi><mi>n</mi></mrow></semantics></math></inline-formula> and <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>F</mi><mn>1</mn></mrow></semantics></math></inline-formula> validation metrics and <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>D</mi><mi>S</mi><mi>C</mi></mrow></semantics></math></inline-formula> test metrics was calculated. Recognition accuracy was analyzed as affected by the optimizer during neural network training. In the context of this problem, neural networks trained using the stochastic gradient descent optimizer with Nesterov momentum were found to have the best generalizing properties. To select the best model during its training on the basis of the validation metrics, the main test metrics of recognition quality (Dice similarity coefficient) were analyzed depending on the validation metrics. The ResNet and DenseNet models were found to achieve the best generalizing properties for our task. The highest recognition accuracy was attained using the U-Net model with a ResNet152 backbone. The results obtained on the test dataset were <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>D</mi><mi>S</mi><mi>C</mi><mo>=</mo><mn>0.9304</mn></mrow></semantics></math></inline-formula> and <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>I</mi><mi>o</mi><mi>U</mi><mo>=</mo><mn>0.9122</mn></mrow></semantics></math></inline-formula>.