As a key material for mechanical manufacturing, the microstructure of spring steel is closely related to its mechanical properties. Traditional identification methods have limited accuracy and are difficult to efficiently and accurately classify and identify microstructures such as tempered flexural, which restricts the optimization of spring steel properties. In this study, a deep residual network and a convolutional neural network are combined to construct a classification and identification model of spring steel tempered flexural microstructure. By establishing the SS- 3000 dataset containing 5 classes of tissues and 600 images of each class, a migration learning strategy was used to optimize the model training process and validate the model performance on the TEST-1000 dataset. In the experimental design, images were normalized to a fixed interval and normalized, and hyperparameter settings such as batch size 32, learning rate 0.0005 and 500 training cycles were used. The results show that the average recognition accuracy of the SE-Resnext101 metallographic tissue recognition model based on migration learning for various types of microstructures of spring steel reaches 98.2%, and the recognition accuracy and recall of tempered flexural reaches 97.82% and 99.04%, respectively, which is significantly better than that of the traditional GLCM+SVM algorithm and other deep learning models. In the comparison experiments, the recognition accuracy of the model for tempered flexural is 5.69%, 7.19% and 5.43% higher than that of VGG19, AlexNet-TL and MobileNetV2-TL, respectively. The study confirms that the combination of deep residual network and convolutional neural network can effectively extract the microstructure features of spring steel, which provides reliable technical support for the study of the relationship between material properties and microstructure.
