MobileTransNeXt: Integrating CNN, transformer, and BiLSTM for image classification

Abstract

Transformers have become popular in computer vision by competing with convolutional neural networks (CNNs). However, CNNs have high potential in deep learning, and models based on transformer and CNN collaboration need to be proposed to achieve better performance. This study presents a new hybrid model that achieves high classification accuracy with fewer learnable parameters, and the main goal is to achieve high classification accuracy with fewer learnable parameters. In this research, first, an innovative deep learning architecture consisting of CNN and transformer collaboration is proposed, and in this model is termed MobileTransNeXt. The recommended MobileTransNeXt is obtained by integrating a transformer into MobileNetV2. MobileTransNeXt creates a feature map using MobileNetV2 and uses a transformer to classify the created feature map. In addition, a MobileTransNeXt-based deep feature engineering (DFE) approach is proposed to take the test classification ability to the next level and show high transfer learning ability. The presented models were tested on two datasets (UC-Merced and NWPU-RESISC45). MobileTransNeXt achieved 96.90 % and 95.18 % accuracy, while the DFE model performed even better, reaching 98.81 % and 95.29 %. These results clearly show that MobileTransNeXt is a new computer vision solution.