Deep Learning-based Landmarks Extraction For Car Frontal View And Its Applications

The design of the front face of the car determines the first impression of the car,which can reflect the distinctive appearance of a brand and also have emotional attributes.It is absolutely an important part of car design.The front face of the car shows various functions and information,and the styling design is thus variegated.Automobile modeling design requires the use of professional CAD software,which is complicated to operate,and very unfriendly to ordinary consumers,and unable to achieve efficient and accurate extraction of car front view feature points.Accurate and efficient extraction of vehicle front view information is the key to achieving accurate styling and designing and perfecting system functions.Therefore,this article discusses the automatic extraction of landmarks on the front face of a car.The realization of fully automatic landmarks extraction greatly reduces the threshold for front face design of automobiles and improves design efficiency,which is of great significance for realizing the forward design of automobiles.However,the complexity of the front face of the car also makes it difficult to realize the definition and extraction of the feature points of the car front face.For this reason,this paper proposes a method of landmarks extraction for front face based on deep learning.Firstly,a database of car front view is created,then a front face classification strategy is formulated,pre-processing of the input car front view is performed,and a method based on convolutional neural network is used to automatically recognize and classify;and then,based on Deep Alignment Network(DAN).This method can detect landmarks in the front view automaticly;further realizes the modeling of front view 2D characteristic curves;and lays a foundation for the completion of more accurate and richer features of 3D characteristic curves modeling and automated surface modeling.This paper did the following research work(1)Creation of front face database,classification library,and front view landmarks tag library.The front view database contains 3657 standard car front view,and corresponding landmarks tag libraries are created.(2)Carry out exhaustive research and summary of the design of the front face of the car.Based on the theory of the gestalt and the positional relationship of the components of the front face,the car front view is divided into six categories,and a car front view classification library is created.An improved front face feature line definition method was proposed,and a uniform representation template was formulated so that each category can effectively reflect the front face's modeling features on the premise of consistent topology.After careful analysis of the structure of the front view of the car,the number and attributes of the landmarks of the front view of the car are defined.(3)Based on the above database,automatic recognition of the front view of the vehicle is performed using a classification and recognition method based on a convolutional neural network.The result is the basis for subsequent feature point positioning.(4)A front-view feature point coordinate database was created,and the DAN algorithm was used to auto-position the front view landmarks.A car front view keypoint coordinate database and a 2D characteristic curves database containing 3657 pictures were created.The landmarks positioning results of the front view of the car can be used as the basis for the subsequent automatic reconstruction of the 2D characteristic curves,which is the core of this paper.From the car front view input to the reconstruction of the front view 2D characteristic curves,the above work completes the automation from input to generation.Through a large number of numerical experiments and error statistics,it is proved that with the method of deep learning,the extraction of the landmarks of the car front face and the automatic reconstruction of the 2D curves model can be completed efficiently and accurately.