Design of Efficient Convolutional Neural Module Based on An Improved Module
Daihui Li, Chengxu Ma, Shangyou Zeng · Advances in Science Technology and Engineering Systems Journal · 2020
In order to further improve the feature extraction performance of the convolutional neural networks, we focus on the selection and reorganization of key features and suspect that simple changes in the pooling layers can cause changes in the performance of neural networks.According to the conjecture, we design a funnel convolution module, which can filter out the key features and perform multi-scale convolution of key features.And we apply this module to the design of high performance small neural networks.The experiments are carried out on 101_food and caltech-256 benchmark datasets.Experiments show that the module has higher performance than classical module, and the small convolution networks based on the module has less parameters and more excellent performance.