SCMA: Exploring Dual-Module Attention With Multi-Scale Kernels for Effective Feature Extraction

Abstract

Feature space enrichment is an integral part of the development of attention mechanisms in Convolutional Neural Networks (CNNs). The ability to efficiently extract channel and spatial information across a variety of scales is crucial. Furthermore, balancing model parameter efficiency while ensuring higher accuracy is a key objective. To create a compelling and robust attention mechanism, channel and spatial attention must be carefully incorporated into CNN architecture. This research work addresses these challenges and presents an attention mechanism called Spatial and Channel aware Multi-scale kernel Attention (SCMA) for CNNs. Our approach leverages the combination of two separate attention modules, one for channel-wise attention and another for spatial attention, in sequential order to refine intermediate feature representations in a CNN. The SCMA module is designed to be compact and universal, capable of being seamlessly integrated into any baseline CNN architecture with minimal parameter overhead, and can be trained in an end-to-end manner. Our empirical findings regarding the utilization of SCMA in conjunction with various CNN architectures for image classification tasks on multiple benchmark datasets including Imagenette, Imagewoof, CIFAR-10, CIFAR-100, and CINIC, affirm the intuition that multi-scale kernels are pivotal for effectively capturing dependencies across both spatial and channel dimensions. In many instances, SCMA exhibits higher performance in terms of accuracy than its state-of-the-art counterparts while keeping the parameter overhead to a minimum.