Research On Precise Workpiece Recognition Algorithms Based On Multi-View Feature Fusion

With the continuous improvement of China’s industrialization level and processing technology,the industrial manufacture also becomes more intelligent in recent years.As an essential part of specific industry,the precision workpiece is required to apply a heat treatment in the processing process,which will take the workpiece off the production line and break the subsequent processing information.To realize the automation of manufacturing,the image processing technology is adopted to identify the types of workpieces before and after the heat treatment,so that the processing information of workpiece is re correlation.The contributions of this work can be summarized as follows:Aiming at the low precision of retrieving the fine-grained workpiece image,a multi view feature fusion image retrieval method for precision workpiece is proposed,namely Double MobileNet-Hash Cosine Similarity(DMN-HCS),and the corresponding algorithms are proposed inspire by the key factors affecting the low accuracy of workpiece image retrieval.1)Aiming at the low retrieval accuracy caused by the small difference of top texture of the fine-grained precision workpiece,we study the top texture features and propose a depthmeasurement-learning-based algorithm for precision workpiece image retrieval,namely Double MobileNet(DMN).Firstly,we improve the original mobilenetv2 to the double output DMN model,and adopt the difference loss function to extract the common and private features of the workpiece image in the same view.Then,the distance between the common features of the top view and the side view is reduced by adopting the similarity loss function to obtain the initial embedding vector.Moreover,the triple central loss function is used to supervise the learning of the initial embedding vector and obtain the embedding vector with strong robustness by reducing the distance within the class and increasing the distance between the classes.Finally,the embedded vector is used to represent the feature coding of top texture features,which realizes the retrieval of workpiece images with top texture differences.The theoretical analysis and experimental results show that the algorithm has stronger representation ability than the conventional feature coding method,and its retrieval accuracy is improved by 7.67%.2)In view of the low retrieval accuracy caused by the small size difference and the side texture difference of the fine-grained precision workpiece,We study the size features and side texture features and improve the hash fingerprint template matching algorithm,and propose a multi view precision workpiece image retrieval algorithm inspired by template matching,namely Hash Cosine Similarity(HCS).Firstly,the dimension features of top view with height view,and the texture feature are all extracted;Then,we calculate the similarity between the test workpiece features with the template library workpiece;Finally,the matching result is obtained by using the similarity distance,which realizes the retrieval of the above workpiece image.Theoretical analysis and experimental results show that this algorithm outperform the traditional template matching algorithm,and the matching accuracy improve 4.82%.Theoretical analysis and experimental results show that the DMN-HCS method proposed in this paper has high retrieval accuracy for workpiece images with top texture difference,top size difference,side height difference and side texture difference,and the retrieval accuracy is 99.30%,which meets the actual production needs of precision workpiece.