Nettop: A light-weight network of orthogonal-plane features for image recognition

Thanh Tuan Nguyen, Thanh Phuong Nguyen

Abstract: In the current light-weight CNN-based networks, convolutional operators are principally utilized to extract feature maps for image representation. However, such conventional operation can lead to lack of informative patterns for the learning process. It is because the operators have just been allocated to convolute on the spatial side of an input tensor. To deal with this deficiency, we propose a competent model to efficiently exploit the full-side features of a tensor. The proposed model is based on three novel concepts as follows. i) A novel grouped-convolutional operator is defined to produce complementary features in consideration of three plane-based volumes that have been correspondingly partitioned subject to three orthogonal planes (TOP) of a given tensor. ii) An effective perceptron block is introduced to take into account the TOP-based operator for orthogonal-plane feature extraction. iii) A light-weight backbone of TOP-based blocks (named NetTOP) is proposed to take advantage of the full-side informative patterns for image representation. Experimental results for image recognition on benchmark datasets have proved the prominent performance of the proposals. The code of NetTOP is available at https://github.com/nttbdrk25/NetTOP.