Research On 3D Model Target Detection Based On Deep Learning

In recent years,3D models have been widely used in industrial manufacturing and other fields for their characteristics of virtualization,digitizationand visualization.Research on the retrieval and recognition of three-dimensional models of mechanical components has become a hot topic.The traditional 3D model recognition retrieval method is to use the feature descriptors to represent the 3D model features,The quality of the recognition results depends on people's subjective understanding of the three-dimensional models andthe feature descriptors' ability about express models.With the development and popularization of artificial intelligence technology,artificial intelligence has appeared in related fields such as industrial manufacturing.The breakthrough development of deep learning has made it widely used in the field of images.Because the appearance property of 3D model is its most intuitive feature expression,this paper uses deep learning technology to deeply study the target detection of 3D model image.In the application scenes of 3D model's identification and location,this paper starts research work based on the Faster R-CNN.The main work contents are as follows:(1)Virtual sample generation and data set production of mechanical components.Firstly,the 3D model of the components is placed in the center position of the sphere space in the virtual environment,the shooting angle is changed by moving the virtual camera,and intercepting images from different angles as samples.Then,according to thecharacteristics of the 3D model image samples,the automatic annotation of the samples is realized,and the data set is completed.The data set contains 21 types of components,and 15,465 images.(2)Fusing multi-layer feature.According to the characteristics of the 3D model image dataset,the shallow convolution feature map of VGG16 is fused with the deep convolution feature map,and the fused resultsare used for target detection.The method makes full use of the underlying information such as edge,shape and position in the image,so that the model detection accuracy reaches 93.75%,the 5.02% is improved,and the test time is shortened by 14 ms.(3)Generating proposal regions based on conv4.In the convolutional neural network,as the number of convolution layers deepens,a large amount of detailed information is lost in the feature map,and the position's information is also reduced.In this paper,the fourth layer convolution feature map is used to realize candidate region generation.The generated candidate region coordinates and the fifth layer feature map are input into the ROIs Pooling layer to realize feature map clipping and coordinate mapping.Compared to the original model,the predicted border and real border IoU coverage increased by 1.88%.(4)Mixed samples achieve real-world scene target detection.In the absence of real samples,this paper proposes to use a large number of virtual samples and a small number of real samples to form a mixed sample training deep learning model,the real samples is tested,the detection accuracy reaches 70.8%,under the premise of ensuring the detection accuracy,the real samples collecting and labeling has been reduced.Finally,the feasibility and effectiveness of the improved method are verified on the 3D model image dataset.Compared with the target detection results,the accuracy of the proposed method in target classification and position prediction has been improved.