Algorithm 1: The implementation of asymmetric multi-weights attention.

Input X: The feature matrix of H × W × C size.

Output X: The resultant matrix of H × W × C size.

(1)Set a 3 × 1 convolution layer and compress the channels to C/2. (2)Use a 1 × 3 convolution layer and expand the channels to C. (3)Calculate spatial size N = H × W − 1. (4)Calculate square D = X − X.mean().pow(2). (5)Calculate channel variance through D/N and derive function F for finding the importance of each pixel as F = D/(4 × (v + lambda)) + 0.5, where lambda is the coefficient value. (6)Adding sigmoid to restrict F. (7)Save the value of the output matrix.