HBONext: HBONet with Flipped Inverted Residual

Abstract

The top-performing deep CNN (DCNN) architectures are presented every year based on their compatibility and performance ability on the embedded edge applications, significantly for image classification. There are many obstacles in making these neural network architectures hardware friendly due to the limited memory, lesser computational resources, and the energy requirements of these devices. The addition of Bottleneck modules has further helped this classification problem, which explores the channel interdependencies, using either depthwise or groupwise convolutional features. The classical inverted residual block, a well-known design methodology, has now gained more attention due to its growing popularity in portable applications. This paper presents a mutated version of Harmonious Bottlenecks (DHbneck) with a Flipped version of Inverted Residual (FIR), which outperforms the existing HBONet architecture by giving the best accuracy value and the miniaturized model size. This FIR block performs identity mapping and spatial transformation at its higher dimensions, unlike the existing concept of inverted residual. The devised architecture is tested and validated using CIFAR-10 public dataset. The baseline HBONet architecture has an accuracy of 80.97% when tested on CIFAR-10 dataset and the model's size is 22 MB. In contrast, the proposed architecture HBONext has an improved validation accuracy of 88.30% with a model reduction to a size of 7.66 MB.