Fig. 3From: Ancient mural segmentation based on a deep separable convolution networkOperational principle of a deep separable network. a Depthwise convolution. b Pointwise convolution. The input is a three-channel color image. The image undergoes depthwise convolution (a) to obtain a feature map (maps*3) on the two-dimensional plane, which further undergoes pointwise convolution (b) to form a feature map (maps*4). This treatment aims to effectively take advantage of different channels to obtain the features at the same spatial positionBack to article page