Deep Residual Networks With a Flask-Like Channel Structure

The development of deep residual network (ResNet) has contributed significantly to the progress of computer vision and image classification, expanding the applicability of convolutional neural networks to different fields. Researchers continue to improve the classification accuracy of ResNet by increasing parameter sizes or model complexity. However, enlarging the network parameter size would significantly increase the training workload. In this work, we adjusted the scale and variation pattern of ResNet18 channel numbers, evaluated their performance differences using different datasets, and designed a flask-like channel structure, which enabled ResNet18 to reduce model parameters while maintaining accuracy. Then, we use MNIST and STL10 datasets validate the effectiveness of FLC structure. Finally, we extend the FLC structure to other ResNet models with different layers, such as ResNet34, ResNet50, ResNet101, and ResNeXt. By testing these ResNet models on the CIFAR10 dataset, our experiments showed that the ResNet models with the FLC structure (namely ResNet_FLC) can maintain or improve the accuracy of the model by approximately 1% while reducing the number of model parameters and FLOPs.