Research On 3D Model Retrieval Technology Based On Freehand Sketch

Three-dimensional models are widely used in the fields of medical research,architectural design,three-dimensional animation production,garden planning,and augmented reality.With the development of society and the continuous improvement of 3D model data acquisition technology,the demand for 3D model products has also increased day by day.The cost of 3D modeling is relatively high.In order to make full use of the existing 3D models,the technology of constructing 3D scenes through information retrieval 3D models has great demand.Among them,the 3D model retrieval technology based on hand-drawn sketches has become one of the research hotspots because of its simple and convenient human-computer interaction.In practical applications,hand-drawn sketches are highly uncertain and abstract,and as 3D models become more and more complex,the retrieval of 3D models based on hand-drawn sketches becomes more and more difficult.In this paper,the data of the 3D model is enhanced,and then based on the deep convolutional neural network method,the global and local features of the image,which are fused as the feature vector for retrieval and matching,are extracted.It avoids the disadvantages of weak traditional feature expression ability,can better adapt to model retrieval in the era of big data,and achieves the purpose of retrieving 3D models based on hand-drawn sketches.The specific content is as follows:(1)In order to match features with hand-drawn sketches,the 3D model is projected into a series of 2D pictures data on which the enhanced processing was carried out.Firstly,three rotations are performed randomly,and each rotation is used to project the eight quadrant directions of the model to enhance the projection perspective of the 3D model,so that the projection perspective of the model could meet the actual requirements.Then,in this paper,the non-maximum suppression and threshold selection of Canny algorithm are further improved,so that the effect of edge detection is enhanced.Finally,the traversal process of the eight-neighborhood contour tracking algorithm is optimized during the contour tracking of the projection view,with which the traversal time and the deviation rate are reduced.After data enhancement,the projection perspectives are convenient for the feature value extraction,which makes the feature extraction model more widely applicable.(2)In order to extract more discriminative feature values,a Feature Extract Joint Learning Network(FEJLN)model that can fuse global and local convolutional features is established.The model uses the residual network as the basic network to extract the low-level features,on which global and local features are extracted respectively and fusesed according to a certain weight to improve the accuracy of model retrieval.The loss functions of the two networks are fused as the total loss function of the FEJLN model,and the parameters of the network are continuously adjusted in the process of finding the minimum value to improve the accuracy and generalization ability of the model retrieval.(3)In the process of extracting local features,a relatively simple and high-performance local feature extraction network(Part Feature Network,PF_Net)is desined.The low-level features extracted by the basic network are horizontally segmented and used as the input of PF_Net.Global average pooling is used instead of the fully connected layer to extract finegrained image features.In order to improve the accuracy of model retrieval,a 1 × 1 convolution layer is added to PF_Net to reduce the dimension of feature vectors,and a batch regularization layer is introduced to prevent the overfitting in the network.(4)Based on the above process,experiments are performed on the SHREC'13 and PSB data sets to verify the effectiveness of the model in this paper.The results show that the accuracy of the retrieval algorithm in this paper is improved more than 10 % compared with other models,which provides a reliable basis for the retrieval of 3D models.