Fu, Yinghua; Liu, Peiyong; Lu, Haiyang; Gupta, Rohit; Zhang, Dawei; (2025) MirageNet: improving the network performance of image understanding with very low FLOPs. Expert Systems with Applications , 286 , Article 128018. 10.1016/j.eswa.2025.128018.

Text Gupta_MirageNet.pdf - Accepted Version Access restricted to UCL open access staff until 15 May 2026. Download (3MB)

Abstract

In order to realize lightweight CNN architectures, an array of methods including model pruning, distillation, and inventive convolutional designs have been developed and employed. However, the computational complexity to approximate 20M FLOPs often leads to significant performance degradation. In this paper, this study presents the innovative Mirage module, designed to produce an expanded array of feature maps through cost-effective processes. By harnessing the power of low-rank matrix factorization and group convolution, this module is capable of generating numerous feature maps, thereby comprehensively exposing the fundamental data and its attributes. MirageNet achieves high accuracy within the range of 18M-21M FLOPs by enhancing inter-group nonlinearity and feature activation and improving the DY-Shift-Max activation function. The incorporation of the Squeeze-and-Excitation (SE) module further optimizes MirageNet's architecture. Experimental results reveal that MirageNet outperforms existing lightweight models, achieving a top-1 accuracy of 63.2 % on ImageNet with 20.7M FLOPs, surpassing MobileNetV3 by 9.2 %.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item