Image Inpainting for Irregular Holes Using Partial Convolutions
Partial Convolutional Layer
只对non-hole的部分做卷积运算
可以理解为对mask做卷积运算,卷积核的权值都为1,相当于对slide window对应的mask小区域求和,也就是公式中的sum(M),如果结果大于0,该小区域卷积之后的结果为1,otherwise为0。这就做到了mask的更新只要网络深度足够大,mask最终都为1。
Network Architecture
## Loss Functions
These are the L 1 losses on the network output for the hole and the non-hole pixels respectively.
The perceptual loss computes the L 1 distances between both I out and I comp and the ground truth, but after projecting these images into higher level feature spaces using an ImageNet-pretrained VGG-16. Ψ p I ∗ is the activation map of the pth selected layer given original input I ∗ . We use layers pool1, pool2 and pool3 for our loss.
We further include the style-loss term, which is similar to the perceptual loss, but we first perform an autocorrelation (Gram matrix) on each feature map before applying the L 1 .
Our final loss term is the total variation (TV) loss L tv : which is the smoothing penalty on R, where R is the region of 1-pixel dilation of the hole region.
Experiments
Cremi数据集中存在几张带黑线的图像,这些黑线是由于切片时,不小心将部分细胞划破导致的,如果不将黑线去除,会影响到细胞分割的效果。
Trainset
Results