Lighter and Faster Cross-Concatenated Multi-Scale Residual Block Based Network for Visual Saliency Prediction

Existing deep architectures for visual saliency prediction face problems like inefficient feature encoding, larger inference times, and a huge number of model parameters. One possible solution is to make the local and global contextual feature extraction computationally less intensive by a novel lighter architecture. In this work, we propose an end-to-end learnable, inter-scale information sharing residual block based architecture for saliency prediction. A series of these blocks are used for efficient multi-scale feature extraction followed by a dilated inception module (DIM) and a novel decoder. We name this network as cross-concatenated multi-scale residual (CMR) block based network, CMRNet. We comprehensively evaluate our architecture on three datasets: SALICON, MIT1003, and MIT300. Experimental results show that our model works at par with other state-of-the-art models. Especially, our model outperforms all the other models with a smaller inference time and a lesser number of model parameters.