Research and Implementation of a Spare Parts 3D Model Retrieval Algorithm Based on the Multi-View Projection Method

Currently, 3D model retrieval has some problems that make it difficult to be applied, such as relying on the initial pose of the model and having unsatisfactory retrieval efficiency and effects. Meanwhile, using deep learning techniques to solve these problems also brings issues like high time cost and poor practicability. In response to this, a new 3D model retrieval framework and algorithm are designed. This algorithm adopts projection technology to generate depth maps from multiple angles at multiple axes positions. Instead of training a model from scratch, it utilizes existing pre-trained models to extract features and then fuses them into new feature vectors. It employs a vector database to achieve efficient retrieval. Based on this retrieval framework and algorithm, experiments have been carried out on the model sets of several mechanical design enterprises. The results show that the new method not only solves the problem of relying on the initial pose of the model but also, compared with deep learning techniques, saves the training time cost. Moreover, it has lower requirements for computer performance, and the retrieval effect is satisfactory. Its practicability has reached a level where it can be put into use. Compared with the deep learning techniques that require several days for retraining, when a new 3D model is added, the new method only takes a few minutes to complete the extraction and storage of the model's feature data.