Feature Extraction For Car Face Geometric Attributes Via Deep Learning

The front face of a car is one of the most important parts of the familiarization of the automobile brand.The stylist conveys the design intention through it,and it is also the first intuitive feeling of the consumers to the automobile.Simultaneously,the urgent need for diversification and individualization makes the front faces complex and changeable.Automobile stylists utilize professional software to design automotive new models,which is complex and tedious.Therefore,extraction of the geometric features for automotive front face efficiently and accurately is the key procedure to optimize the process of automotive design.In addition,precise geometric feature extraction and annotation of automobile images can be widely used in fine-grained classification of automobile models,extraction of automobile components,state detection and other scenarios.For this reason,this paper proposes an automatic extraction method of multi-class geometric features of automobile front face based on deep learning,which not only improves the efficiency of design phase,but also has great significance in the fields of automatic driving and intelligent transportation.This paper has done the following research work:· Creation of a unified car front face annotation data set Auto Morpher/FFD-Cars22,which is suitable for vehicle front face component detection,semantic segmentation,key point localization and other issues.Our database contains 4,457 real-life pictures of the car,covering 22 brands which are common in the Chinese market.· A component detection and semantic segmentation method for the front face of a car is proposed.This paper proposes the use of the Refine Det detection method to detect the front part of the car.Because it uses a two-stage detection idea and feature fusion operation,the accuracy far exceeds YOLO and SSD without the occurrence of missed detection and repeated detection.Furthermore,the Deep Lab algorithm is adapted to semantically segment the front face of the car.Owing to the random condition field and hole convolution used by Deep Lab,compared with FCN and Mask-RCNN,the proposed algorithm can achieve efficient car front face semantic segmentation.· Two efficient key point extraction methods for the front face of the car are proposed.First of all,this paper proposes DAN-Dense Net.Due to its deeper network structure and more complex feature representation,experiments show that its accuracy can be improved by 6.4% compared with the original network.Furthermore,this paper proposes the DAN-Mobile Net algorithm,which can reduce the running time of the original network by 8% while maintaining the original network with 99% accuracy.In summary,benefiting from deep learning,this paper realizes the high efficiency and accurate extraction and segmentation of various car front face geometric features,and we had created a large-scale car front face labeling database Auto Morpher/FFD-Cars22.