A spatial attention map is subsequently derived by concatenating channel-wise maximum-pooled and average-pooled statistics, passing them through an additional one-by-one convolution, and applying a sigmoid activation function.
Concurrently, a complementary spatial attention map is generated by passing channel-pooled statistics through a large seven-by-seven spatial convolution kernel.
Complementary sensitivity analyses verified the optimal hyperparameter configurations governing spatial masking granularity, attention dimensionality, and classifier regularization.
Subject of Research: Deep learning architecture for handicraft pattern and surface texture recognition Article Title: Handicraft pattern design recognition and classification based on SMPANet Article References: Niu, F., Xie, T., & Ma, S. (2026).
Handicraft pattern design recognition and classification based on SMPANet.