Predicting visual saliency via a dilated inception module-based model
Sheng Chih Yang, Weisi Lin · International Workshop on Advanced Image Technology (IWAIT) 2019 · 2019
With the advent of deep convolutional neural networks (DCNN), the improvements in visual saliency prediction research are impressive. Despite this, it is still needed to fully characterize the multi-scale saliency-influential factors into the current deep saliency framework for further improvement. However, the existing approaches aiming at capturing multi-scale contextual features either suffer from the heavy computation or limited performance gain. To overcome this, a lightweight yet powerful module for fully exploiting multi-scale contextual features is desired. In this paper, we propose a DCNN-based visual saliency prediction model to approach this goal. Our model is inspired by the GoogleNet, which use the inception module to capture multi-scale contextual features at various receptive fields. Specifically, we revise the original inception module to have more powerful multi-scale feature extraction capacity and less computation load by utilizing dilated convolutions to replace the original standard ones. The whole model is trained end-to-end and is efficient to achieve real-time performance. Experimental results on several challenging saliency benchmark datasets, including SALICON, MIT1003, and MIT300, demonstrate that our proposed saliency model can achieve state-of-the-art performance with competitive inference time.