基于特征稀疏化的粉尘图像深度预测

Dust Image Depth Prediction Based on Feature Sparsity

【目的】针对粉尘环境中单幅图像深度预测精度低的问题,提出了一种基于输入特征稀疏化的粉尘图像深度预测网络。【方法】使用粉尘图像的直接传输率与深度的关系设计预估计深度网络,利用图像颜色衰减先验原理进一步获取粉尘图像的稀疏深度特征。将该稀疏深度特征与粉尘图像一起作为深度预测网络的输入。深度预测网络以“编码器-解码器”为模型框架,编码器中使用残差网络(ResNet)对粉尘图像进行编码,设计融合通道注意力机制的稀疏卷积网络对稀疏深度特征进行编码。解码器中采用反卷积以及多尺度上采样的方法,以更好的重建稠密的深度信息。使用最小绝对值损失和结构相似性损失作为边缘保持损失函数。【结论】在NYU-Depth-v2数据集上的实验结果表明该方法能够从粉尘图像中有效预测深度信息,平均相对误差降低到0.054,均方根误差降低到0.610,在δ<1.25时准确率达到0.967.

【Purposes】Aiming at the problem of low accuracy of single image depth prediction in dusty environment,a dust image depth prediction network based on sparse input features is proposed.【Methods】First,by using the relationship between the direct transmission rate of dust image and depth information,a depth prediction network is designed to obtain a depth prediction map.With the prior principle of image color attenuation,the sparse depth features of the dust image are further obtained from the estimated depth map.Then,the sparse depth features and dust images are used as the input of the depth prediction network.The deep prediction network uses an“encoder-decoder”model framework,and the two inputs in the encoder are encoded by different networks.Among them,the dust image is encoded by the residual network,which solves the problem of gradient dispersion and accuracy degradation in the deep network.This pre-diction method not only ensures the accuracy but also controls the speed.The sparse depth fea-ture adopts a sparse convolutional network with a fusion channel attention mechanism.Coding makes effective feature maps weight more and invalid or less effective feature maps less.【Conclusions】The resulting output is then channel-fused and fed into the decoder.The decoder uses de-convolution and multi-scale upsampling to design an upsampling module,and uses bicubic up-sampling to better reconstruct dense depth information.Minimum absolute value loss and struc-tural similarity loss are adopted as edge preserving loss functions.The experimental results on the NYU-Depth-v2 dataset show that the method can effectively predict the depth information of dust images,the average relative error is reduced to 0.054,the root mean square error is reduced to 0.610,and the accuracy ofδ<1.25 is 0.967.