Abstract:

In view of the problem of insufficient feature utilization of low-resolution image in super-resolution reconstruction, a multi-feature gated feedback residual network is proposed based on feedback mechanism and attention mechanism. The network has a simple structure and realizes the reuse of network parameters with a circular way, which can save compute resource effectively. The output features of network iteration are retained to achieve multi-feature fusion. In addition, a further feature refine block is used to extract the reconstructed high-resolution image features to obtain better reconstruction result. Experimental results on five test data sets show that when the scale factor is 4, the peak signal-to-noise ratio of the proposed network is 32.50 dB, 28.83 dB, 27.75 dB, 26.65 dB and 31.12 dB, respectively. Compared with the comparison networks, the test results of the proposed algorithm are significantly improved.

Key words: convolutional neural network, residual network, feedback mechanism, attention mechanism, gated unit, multi-feature, feature extraction, super-resolution reconstruction