Visual Saliency Prediction Using a Mixture of Deep Neural Networks.

Visual saliency models have recently begun to incorporate deep learning to achieve predictive capacity much greater than previous unsupervised methods. However, most existing models predict saliency without explicit knowledge of global scene semantic information. We propose a model (MxSalNet) that incorporates global scene semantic information in addition to local information gathered by a convolutional neural network. Our model is formulated as a mixture of experts. Each expert network is trained to predict saliency for a set of closely related images. The final saliency map is computed as a weighted mixture of the expert networks' output, with weights determined by a separate gating network. This gating network is guided by global scene information to predict weights. The expert networks and the gating network are trained simultaneously in an end-toend manner. We show that our mixture formulation leads to improvement in performance over an otherwise identical nonmixture model that does not incorporate global scene information. Additionally, we show that our model achieves better performance than several other visual saliency models.

Quality Robust Mixtures of Deep Neural Networks.

We study deep neural networks for classification of images with quality distortions. Deep network performance on poor quality images can be greatly improved if the network is fine-tuned with distorted data. However, it is difficult for a single fine-tuned network to perform well across multiple distortion types. We propose a mixture of experts based ensemble method, MixQualNet, that is robust to multiple different types of distortions. The "experts" in our model are trained on a particular type of distortion. The output of the model is a weighted sum of the expert models, where the weights are determined by a separate gating network. The gating network is trained to predict weights for a particular distortion type and level. During testing, the network is blind to the distortion level and type, yet can still assign appropriate weights to the expert models. In order to reduce the computational complexity, we introduce weight sharing into the MixQualNet. We utilize the TreeNet weight sharing architecture as well as introduce the Inverted TreeNet architecture. While both weight sharing architectures reduce memory requirements, our proposed Inverted TreeNet also achieves improved accuracy.