BAM: a balanced attention mechanism to optimize single image super-resolution

Recovering texture information from the aliasing regions has always been a major challenge for single image super-resolution (SISR) task. These are often submerged in noise so that we have to restore details while suppressing noise. To address this issue, propose an efficient Balanced Attention Mechanism (BAM), which consists of Avgpool Channel Module (ACAM) and Maxpool Spatial (MSAM) parallel. ACAM is designed suppress extreme large-scale feature maps, MSAM preserves high-frequency details. Thanks parallel structure, these two modules not only conduct self-optimization, but also mutual optimization obtain balance reduction restoration during back propagation process, structure makes inference faster. verify effectiveness robustness BAM, applied it 10 state-of-the-art SISR networks. The results demonstrate BAM can efficiently improve networks' performance, those originally with attention mechanism, substitution further reduces amount parameters increases speed. Information multi-distillation network (IMDN), representative lightweight attention, when input size 200 × 200, FPS proposed IMDN-BAM precedes IMDN {8.1%, 8.7%, 8.8%} under three SR magnifications 2, 3, 4, respectively. Densely residual Laplacian (DRLN), heavyweight scale 60 60, DRLN-BAM {11.0%, 8.8%, 10.1%} faster than DRLN 4. Moreover, present dataset rich real scenes, named realSR7. Experiments prove achieves better on area.