Research On 3D Interior Model Retrieval Based On Visual Feature

With the rapid development of computer graphics technology,3D interior models have been widely used in interior design industry.Designers can design and modify interior space efficiently by simply dragging and replacing 3D interior models in interior design software,which greatly improves design efficiency and reduces design cost.At present,most interior design softwares provide lots of 3D interior models for users to select.However,in the process of selecting models,users can only use limited keywords to retrieve in the form of text,and can not accurately obtain the desired model,which has great limitations.Therefore,a 3D interior model retrieval method based on visual feature is studied in this paper.This method uses image as the retrieval input source,transforms the retrieval of 3D model into the retrieval of 2D image,which can help users find the target model quickly and accurately.The main research contents of this paper include:(1)Kinds of interference factors in the images uploaded by users are analyzed,and it is proposed that before feature extraction,the main body of the image including the model should be segmented from the original image.In this paper,the corresponding data sets are used to compare the actual performance of various mainstream object detection algorithms,and Faster R-CNN is selected to detect the main body of the image.On this basis,the residual network and feature pyramid network are used to improve the detection accuracy of the algorithm.(2)Aiming at the problem that the traditional feature extraction network can't realize the image retrieval of case level,a feature extraction network based on metric learning is designed.By constructing triplet samples online and using triplet loss function to train the network,the network can distinguish the samples in a fine-grained way.In addition,the spatial pyramid pooling layer is added to the network to fuse the multi-scale features in the image and improve the robustness of the network to scale changes.(3)In order to solve the problem that the feature extraction network can't be applied to large-scale model retrieval because of the high dimension of the output feature vector,a deep hashing network which introduces segmented coding and binary constraint term is designed.The network uses the segmented coding structure to reduce the redundant information in the output hash code,and by adding a binary constraint term to the loss function,the output distribution of the network is constrained and the quantization error of the features is reduced.This network can compress the high-dimensional feature vector into the low-dimensional hash code,ensure a certain retrieval accuracy,and greatly improve the retrieval speed at the same time.(4)On the basis of the above research,this paper integrates all links of 3D interior model retrieval,designs the corresponding retrieval process,and realizes the retrieval system in the form of B/S framework by using the front and back end technology of web.This paper tests the retrieval performance of the system through experiments,and analyzes and evaluates the retrieval results.The experiment results show that the system has a good performance in the retrieval accuracy and speed,and is competent for large-scale model retrieval.