Feature Diversity Learning for a Compact CNN Module

Designing a compact network with minimal performance degradation offers a promising prospect for limited computational resources. To this end, existing methods commonly focus on simply reducing redundant features from the perspective of feature redundancy rather than feature diversity. In fact, this way does not mean an increase in feature diversity, thereby obtaining a limited performance improvement. To address this issue, a <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</b>imilarity attention and <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</b>erial <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</b>iversity <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</b>onvolution, named SSDConv, for a compact module by exploring the diversity features from the inter-channel similarity and the structure of the filter deployment. Specifically, feature diversity is enhanced by guiding filters to learn diversity features through feature similarity attention and designing a serial convolution structure to extract multi-scale features, respectively. A series of experiments demonstrate that the proposed SSDConv, instead of standard convolutions in various CNN architectures, significantly reduces FLOPs and parameters while achieving superior accuracy on CIFAR-10 (92.67%) and ImageNet (77.07%). It also delivers competitive performance in downstream tasks, improving mAP by 39.3% on COCO and increasing mIoU by 3.2% and mAcc by 1.4% on CityScapes. More importantly, the proposed SSDConv can readily replace the existing convolution modules as a plug-and-play component, and it is promising for extending to a wider range of scenarios.