Research On 3D Model Retrieval Technology Based On Sketch

In recent years,with the continuous development of three-dimensional(3D)technology,3D models have gradually come into peoples' sights,and play important roles in traditional manufacturing,mechanical automation,autonomous driving,virtual reality(VR)and other fields.When retrieving information,the public does not only rely on traditional methods,such as text and image,etc.As one of main carriers of visual information,3D models have gradually attracted attention.Therefore,how to find the ideal target model in a large amount of data has become a hot topic in the field of computer vision.Sketch is an intuitive representation of human brain consciousness.So,the retrieval method of 3D model based on sketch is simple and can express specific needs of the user clearly.This paper mainly studies 3D model retrieval technology based on sketch.The interactive attention module is used to optimize the convolutional neural network(CNN)for extract more effective features.The traditional feature similarity and depth feature similarity were combined to further optimize the results of sketch based 3D model retrieval.Main research contents of this paper are as follows:First,a two-stage projection method is proposed to improve the quality of the projection views.The latitude with the highest predicted probability is selected as the optimal viewing angle,the three-dimensional model is re-projected on this viewing angle to obtain multiple views,and each projected view is preprocessed.The principle of CNN is studied.The mixed view set is used to train CNN.Finally,the optimized network is applied to extract features of sketch and the projection view for similarity calculation.Secondly,an interactive attention mechanism is proposed to optimize original CNN.The optimized network is adopted to extract deep features and similarity calculation is performed.Finally,the principle of 3D shape distribution features is studied,and an improved 2D shape distribution feature is proposed.2D shape distribution feature similarity between the sketch and the projection view,Gist feature similarity and depth feature similarity are weighted and merged.Finally,similarity calculation is performed.Experimental results show that compared with traditional feature extraction algorithm and CNN feature extraction one,the multi-feature weighted fusion method proposed in this paper can improve the retrieval effect to a certain extent.