MSF-ECA: Multi-Scale Fusion Efficient Channel Attention with Residual Gating for Lightweight Classification

Lightweight convolutional neural networks often lack adequate channel-wise modeling capacity for small-scale image classification. We propose MSF-ECA (Multi-Scale Fusion Efficient Channel Attention), a lightweight module that captures multi-scale channel interactions with minimal overhead. The method uses global average pooling to generate channel descriptors, processes them through parallel 1D convolutions of varying kernel sizes, and fuses the responses using a learnable vector α and sigmoid activation. A trainable residual gate β enhances stability and preserves pre-trained features. The module adds only ~40 parameters and integrates seamlessly into intermediate layers (e.g., layer2, layer3) of architectures like ResNet-18. On CIFAR-10, MSF-ECA-ResNet18 outperforms baseline and ECA variants with NO additional FLOPs cost, providing improved representational capacity with minimal computational increase.