Improved Dual-Channel Attention Mechanism Image Classification Method for Lightweight Network

Abstract

In order to solve the problems of large volume and high hardware requirements of deep convolutional neural network model in missile-borne terminal environment, a lightweight network structure based on improved dual-channel attention mechanism is constructed. Aiming at the problem that the network lightweight will sacrifice the classification accuracy, taking the MobileNetV2 lightweight network as the network basic structure, the self-designed attention module is introduced, and the MobileNetV2 network architecture based on SPP-DCAM module is designed to increase the weight of the significant feature map of convolutional layer learning, so as to improve the classification accuracy. The designed spatial information is input in parallel with channel information. By defining 1×1 and 3×3 small convolution, the computation amount and computational complexity are reduced on the basis of ensuring the lightweight of the structure. Finally, an experimental comparison is conducted on the cifar-100 image classification dataset. The results show that the classification accuracy of improved MobileNetV2 is better than that of the traditional VGG16, ResNet18 and DenseNet convoluted networks with the same number of parameters and computational complexity, and the comprehensive performance is stronger, which is more suitable for rapid classification under the condition of limited onboard computational resources.

Keywords

• |mobilenetv2|attention mechanism|image classification|convolutional neural network|lightweight network|artificial intelligence