Saliency Prediction via Multi-Level Features and Deep Supervision for Children with Autism Spectrum Disorder

This paper proposes a novel saliency prediction model for children with autism spectrum disorder (ASD). Based on the convolutional neural network, the multi-level features are extracted and integrated to three attention maps, which are used to generate the predicted saliency map. The deep supervision on the attention maps is exploited to build connections between ground truths and the deep layers in the neural network during training. Furthermore, by performing the single-side clipping operation on the ground truths, our model is encouraged to enhance the capacity of better predicting the most salient regions in images. Experimental results on an ASD eye-tracking dataset demonstrate that our model achieves the better saliency prediction performance for children with ASD.