改进的神经渲染方法在建筑施工场景中的应用

Application of Neural Rendering Based Visual Synthesis in Construction Scene

针对神经辐射场(NeRF)应用于室外施工场景时,由于室外场景中光照难以捕捉和施工场景的前景背景范围差异过大,在新颖视图中会出现模糊和伪影等现象,通过分析提出改进的视觉合成方法。首先通过SFM算法从RGB图像中获得相机参数,实现对室外施工场景的表示;接着引入预训练编码器中生成的向量,并加入渲染网络中减少光照的影响;最终将图像中的前景与背景分离开进行体绘制渲染,从而提高视觉合成的效果。基于室外施工场景数据集,与3种方法进行比较,结果表明,所提方法在峰值信噪比(PSNR)和结构相似性(SSIM)中分别较其中最优方法提高12.2%与10.9%。整体看来,所提方法在室外施工场景中生成的新颖视图具有更好的细腻度。

When neural radiation field(NeRF)is applied to outdoor construction scenes,due to the difficulty in capturing the lighting in outdoor scenes and the large difference between the foreground and background ranges of construction scenes,blurring and artifacts will appear in novel views.Through analysis,an improved visual synthesis method is proposed.First,camera parameters are obtained from RGB images through SFM algorithm to represent outdoor construction scenes.Then,the vector generated from the pre training encoder is introduced and added to the rendering network to reduce the impact of light.Finally,the foreground and background in the image are separated for volume rendering,which improves the effect of visual synthesis.Based on the outdoor construction scene data set,compared with three methods,the results show that the proposed method improves the peak signal-to-noise ratio and structural similarity by 12.2% and 10.9% respectively compared with the best method.It appears that the novel views generated by the proposed method in outdoor construction scenes have better finesse.