GSNet: Group Sequential Learning for Image Recognition

Abstract

In recent years, deep learning has achieved great successes in the field of image cognitive learning, and designing a well-behaved convolutional neural network (CNN)-based architecture has become a challenging and important problem. The traditional group convolution cannot effectively address the severe problem of “information blocking”; hence, this work proposes an efficient CNN-based model to achieve an effective exchange of information between channels. A novel Group Sequential (GS) learning that uses a channel split operation and sequential learning methods is introduced to improve the recognition performance by increasing information communication. Several state-of-the-art models are developed based on GS blocks, and these blocks significantly boost the performance and robustness of the CNN-based models. Extensive experiments are carried out to evaluate the promising performance of the proposed GSNet framework, and experimental results show its superiority on several benchmarks (i.e., the CIFAR-10, CIFAR-100, Tiny ImageNet, ImageNet, and FOOD-101 dataset). Moreover, compared with traditional residual networks (e.g., ResNet-101), the proposed network has achieved a great improvement with fewer parameters, and the error rate of models on the FOOD-101 dataset decreases from 19.08 to 16.02%. The proposed GS block method has significant potential to improve the performance for image recognition, and advance the development of cognitive computation. The results demonstrate the superiority of the proposed method and indicate excellent generalization ability. Code is available at: https://github.com/shao15xiang/GSNet .