A nonlinear feature fusion by variadic neural network in saliency-based visual attention

This study presents a novel combinational visual attention system which applies both bottom-up and topdown information. This can be employed in further processing such as object detection and recognition purpose. This biologically-plausible model uses nonlinear fusion of feature maps instead of simple superposition by employing a specific Artificial Neural Network (ANN) as combination operator. After extracting 42 feature maps by Itti's model, they are weighed purposefully through several training images with their corresponding target masks to highlight the target in the final saliency map. In fact, the weights of 42 feature maps are proportional to their influence on finding target in the final saliency map. The lack of bottom-up information is compensated by applying top-down information with available target masks. Our model could automatically detect the conceptual features of desired object only by considering the target information. We have tried to model the process of combining 42 feature maps to form saliency map by applying the neural network which resembles biological neural network. The Experimental results and comparing our model with the basic saliency model using 32 images of test dataset indicate a noticeable improvement in finding target in the first hit.