Multi-scale feature fusion filtering module in convolutional neural networks

Recent research has demonstrated that the multiscale representation of CNNs is considerably improved by the attention mechanism. However, the majority of current multiscale feature representation techniques only employ a small number of attention blocks in the attention mechanism, ignoring multiscale-level contextual data. This research suggests a multiscale fused feature filtering module as a solution to this issue (MFFFM). It enables branch-specific feature-selective learning of multiscale contextual data. These branches employ various sizes of null convolutions, and they further employ spatial injection correlation and channel correlation to provide channel feature responses that are adaptive. Tiny ImageNet, CIFAR-100, and MS COCO dataset experimental findings demonstrate that MFFFM provides extremely competitive outcomes in comparison to earlier baseline models.