Research And Implementation Of 3D Reconstruction Technology Of Monocular And Multi View Clothing Image

With the development of mobile Internet technology and virtual reality technology,3D clothing model is widely used in the fields of clothing design,personalized customization,virtual fitting,film and television entertainment,animation and game.The research of 3D clothing model reconstruction attracts more and more scholars in related fields such as clothing,graphics and so on.Due to the soft material,easy to deform,easy to fold and complex structure,the existing 3D garment model reconstruction needs high cost and low efficiency.In order to overcome the shortcomings of existing clothing modeling methods,this paper realizes the 3D reconstruction of monocular multi view clothing image,and studies the key technologies in the process of 3D reconstruction of clothing model.The specific research contents are as follows:1.Target recognition and instance segmentation of multi view clothing image sequence.Because of the low resolution and noise,the multi view clothing image sequence acquired by common image acquisition equipment will reduce the extraction and registration accuracy of image feature points,thus affecting the reconstruction effect.In order to solve these problems,after preprocessing the image sequence,such as noise reduction,we use mask r-cnn to recognize the clothing image sequence and segment the example,which greatly reduces the influence of redundant noise outside the object.2.Feature point extraction and matching of clothing image.In this paper,SIFT algorithm is used to extract the feature points and matching point pairs of garment image.In the process of feature point matching,SIFT algorithm ignores the correlation of space vector,which will produce more mismatches.Aiming at the problem of high mismatches rate when ransanc algorithm eliminates the mismatches of SIFT algorithm,cosine similarity algorithm is used to filter the mismatches.At the same time,ran algorithm is used to filter the mismatches Some improvements have been made in the long time and low efficiency of Sanc operation,which further improves the accuracy of feature point matching.3.Reconstruction of 3D clothing model.In this paper,the motion recovery structure algorithm(SFM)is studied.The sparse point cloud data of clothing is obtained by the feature point matching algorithm proposed in this paper.The sparse point cloud data of clothing is transformed into dense point cloud data by CMVs / PMVs method.Finally,the 3D model of clothing is obtained by Poisson surface reconstruction.The experimental result show that the method proposed in the paper can effectively reduce the noise of point clouds,thus reducing the mismatch rate in feature matching;in the case of providing the same number of image sequences,compared with the Structure From Mothion reconstruction method,the reconstruction time-consuming is less,which can well represent the 3D surface details of clothing.The monocular vision device used in this paper is simple,low cost,less scene limitation and high robustness,which is suitable for the reconstruction of 3D garment model in batch.