A casting defect detection algorithm based on multi-dimensional efficient feature fusion convolution

In the sand casting process, identifying and sorting defects such as surface porosity, inclusions, and cracks are crucial for ensuring casting quality. Existing casting defect detection algorithms are deficient in recognition accuracy and detection efficiency. To solve these problems, this study proposes an improved convolutional layer architecture based on the YOLOv8 model: efficient feature fusion convolution (EFFC) and multi-dimensional efficient feature fusion convolution (MEFFC). These architectures can simultaneously capture multi-scale features of casting defects while maintaining high computational efficiency. In addition, to improve coverage when detecting minor target defects, it is proposed to extend the pooling layer into a large kernel convolutional layer, effectively enhancing detection performance. By testing a self-constructed casting defect dataset, it is found that the newly proposed C2f-MEFFC and C2f-EFFC models improve the mean average precision by approximately 5.65% and 4.98%, respectively, compared to the original YOLOv8 model. Additionally, the number of model parameters for C2f-EFFC is reduced by about 9.76%. Introducing a large kernel convolutional (spatial pyramid pooling fast) SPPF improves the mean average precision of the model by 5.9% over the original YOLOv8 model. Ultimately, the YOLO-MEFFC neural network structure shows higher detection accuracy and efficiency for minor casting defects, achieving a maximum increase of 8.4% in the mean average precision compared to the original YOLOv8 model. The experimental results demonstrate that the newly proposed model outperforms existing mainstream models in terms of overall performance and significantly enhances the detection accuracy of casting defects.