Spatial and Channel Attention Integration with Separable Squeeze-and-Excitation Networks for Image Classifications

Abstract

In recent years, convolutional neural networks (CNNs) have performed remarkably well in various computer vision tasks. Attention mechanisms have been extensively explored to enhance the discriminative and representational ability of CNNs. Among these, Squeeze-and-Excitation (SE) networks have shown significant effectiveness by adaptively recalibrating feature maps. This paper proposes a novel architecture by integrating spatial and channel attention mechanisms using separable SE (SC-SE) Layers. Our proposed SC-SE layer with 1D CNN block is applied to the SqueezeNext architecture to construct our SC-SE network (SC-SENet). The proposed SC-SENet is designed to effectively capture spatial and channel-wise dependencies within feature maps. By decoupling spatial and channel attention modules, our model achieves efficient computation and alleviates overfitting. Furthermore, we introduce a novel separable strategy in the SE block, enabling effective feature re-calibration across space with reduced computational cost. Experimental results on benchmark datasets demonstrate the superiority of our proposed SC-SENet over the state-of-the-art SENet methods in terms of accuracy and efficiency. The extensive experiment shows that our proposed SC-SENet for $44$ layers achieved $99.76\%$, $99.56\%$, $96.98\%$, and $99.78\%$ for $0.58M$ trainable parameters and $6.2M$ FLOPS on 2 and 3 classes lung image segmentation, and ``original'' and ``modified'' malaria parasites detection datasets. It also shows state-of-the-art performance on these datasets. We also experimented with some image classification datasets where our proposed SC-SENet performed better than related recently published works.