The impact of surface defects in castings on product quality and production safety was introduced, emphasizing the importance of precise identification and classification of defects. Traditional manual detection methods have problems such as low efficiency and instability. In recent years, the application of deep learning algorithms in the field of defect detection has gradually received attention. The article reviews the deep convolutional neural network model constructed by Jia Minping et al., the improved AlexNet proposed by Zhao Peng et al., the cascaded autoencoder structure proposed by YUN JP et al., the convolutional denoising autoencoder based on fully connected layers proposed by Luo Yuetong et al., and the improvements made to the YOLOv3 model by WANG L and DUAN L. These studies have made some progress in improving detection accuracy and efficiency, but there are still challenges such as a small sample size, multiple parameters, and large computational complexity. This study optimized the algorithm model through data preprocessing, sample image expansion, and Mosaic data augmentation to improve the comprehensive detection effect of surface defects in castings.

1. Dataset preprocessing work

The preprocessing of the dataset for the surface defect detection model of castings was introduced, including sample size expansion and image enhancement techniques. To solve the problem of overfitting caused by insufficient sample size, the sample image is expanded through operations such as flipping, rotating, and color grading to improve the robustness and generalization ability of the model. At the same time, using Mosaic data augmentation methods, especially 9-Mosaic technology, multiple images are randomly selected for cropping, rotation, and stitching, effectively enriching sample diversity, increasing the number of small target samples, simulating complex environments, reducing the model's hardware performance requirements, and improving training effectiveness.

Research on Improving YOLOv5 Model Method

We conducted in-depth research on the improvement method of the casting surface defect detection model based on CCD-YOLOv5 algorithm. Firstly, the native YOLOv5 detection model is introduced, which is a single-stage object detection algorithm that generates target candidate regions through grid partitioning and performs classification and localization simultaneously in the neural network. The YOLOv5 series models have improved detection efficiency and accuracy through various improvements, such as increasing image resolution and introducing data augmentation. YOLOv5s was specifically chosen as the base model due to its lightweight and wide range of applications.

Surface defect detection test for castings

Introduced the experimental environment, experimental indicators, and ablation test of the casting surface defect detection model based on CCD-YOLOv5 algorithm. The experimental environment configuration includes Python, Pycharm, Pytorch framework, as well as dataset preparation and enhancement. We used cross validation for model training, set parameters such as input image size, batch size, optimizer, and introduced an early stopping strategy. The experimental indicators include accuracy, recall, and class average accuracy (mAP), and the model performance is evaluated through P-R curves and AP values. In the ablation experiment, the performance improvement of the improved model was verified by introducing C2f module, CA attention mechanism module, and decoupling head module. The improved CCD-YOLOv5 model has shown improvements in accuracy, recall, and mAP values, but its recognition performance for certain defect types still needs to be optimized. The analysis of the accuracy recall curve shows that the model performs well in feature extraction and target localization, but there are still challenges when dealing with multiple types and small target features.

4 Conclusion

A casting surface defect detection method based on improved YOLOv5 was proposed, which improves sample richness through data augmentation and 9-Mosaic enhancement strategy, improves model structure to achieve lightweight and gradient information enrichment, introduces CA attention mechanism to reduce redundant information interference, enhances feature expression, replaces coupling head module with decoupling head module to enhance global information acquisition. The improved model performs better in detection accuracy and efficiency.