Joint operation and attention block search for lightweight image restoration

Highlights：

• Based on the differentiable architecture search, we propose a joint operation and attention block search algorithm that enables operation type search and attention mechanism search simultaneously. It provides more room for searching for better networks of image restoration.

• We propose the operation block search space and attention block search space to find the optimal combination of operation block and attention block, respectively. Taking these two searched modules together, we construct a joint search module to formulate the final network.

• We propose the cross-scale fusion module (CSFM) based on butterfly structures and multi-scale features, which can be embedded in the searched network, thus helping to expand representation space for achieving more powerful networks.