Graph Matching Method for Complex Nesting Scenes Based on Multimodal Mask Features and SplineCNN

In the production process of steel plate cutting and sorting in shipyards, it is the key link to improve the efficiency of automated production to achieve accurate matching of the contours of the cut parts in real scenarios and the contours of CAD nesting drawings. However, in the real production environment, there are serious physical disturbances such as uneven lighting, cutting waste bonding, and sheet mutual occlusion. These complex working conditions make the traditional single feature extraction effect poor. Therefore, this paper proposes a matching architecture based on dynamic fusion of multimodal mask features and continuous spatial graph convolutional network, extracts contour features from three dimensions: deep semantics, geometric constraints and edge shape, realizes adaptive feature fusion through layer normalized latent space projection mechanism, and finally uses graph neural network to aggregate the geometric invariant features of the global spatial layout and calculates the similarity matrix to obtain the contour matching results. The comparison and ablation experiments on the real typesetting dataset show that the proposed method shows excellent robustness and matching accuracy, and the optimal matching accuracy of 70.79% is achieved under strong interference. In summary, this paper provides a new algorithm with great potential for plate identification and sorting under severe working conditions.