An accelerating convolutional neural networks via a 2D entropy based-adaptive filter search method for image recognition

The success of CNNs for various vision tasks has been accompanied by a significant increase in required FLOPs and parameter quantities, which has impeded the deployment of CNNs on devices with limited computing resources and power budgets. Network pruning, which compresses and accelerates CNN models, is an effective solution to this issue. Some studies have considered pruning as a special case of neural network search (NAS) in recent years. However, existing techniques are often computationally complex or prone to sub-optimal pruning results. As such, this paper proposes a novel acceleration method via a 2D Entropy based-Adaptive Filter Search (2EAFS). The importance of corresponding filters, measured by utilizing the amount of information contained in feature maps, is employed as a theoretical guide to simplify the complex exhaustive search process. Information entropy is then normalized layer by layer and the resulting value is used to calculate a layer-wise importance score in a single step. Additionally, a sparse constraint equation is constructed based on the negative correlation between filter pruning rates and the importance of convolutional layers. The Nelder–Mead search algorithm is then adopted to quickly and adaptively determine the optimal pruning architecture. Finally, importance weights are inherited using the pruning rate and 2D entropy and model performance are restored through fine-tuning. Extensive experiments conducted with the CIFAR-10/100, ILSVRC-2012, NWPU-RESISC45 and CUB-200-2011 datasets showed this approach achieved considerable accuracy increases, with significant reductions in FLOPs and required parameters that surpassed current state-of-the-art methods by a wide margin. For example, 2EAFS achieved a 44.1% reduction in FLOPs over ResNet-50, with only a 0.53% Top-5 accuracy decrease for ILSVRC-2012.