Reproduction of Lateral Inhibition-Inspired Convolutional Neural Network for Visual Attention and Saliency Detection
Authors:
Filip Marcinek
Abstract:
In recent years, neural networks have continued to flourish, achieving high efficiency in detecting relevant objects in photos or simply recognizing (classifying) these objects - mainly using CNN networks. Current solutions, however, are far from ideal, because it often turns out that network can be effectively confused with even natural images examples. I suspect that the classification of an obj…
▽ More
In recent years, neural networks have continued to flourish, achieving high efficiency in detecting relevant objects in photos or simply recognizing (classifying) these objects - mainly using CNN networks. Current solutions, however, are far from ideal, because it often turns out that network can be effectively confused with even natural images examples. I suspect that the classification of an object is strongly influenced by the background pixels on which the object is located. In my work, I analyze the above problem using for this purpose saliency maps created by the LICNN network. They are designed to suppress the neurons surrounding the examined object and, consequently, reduce the contribution of background pixels to the classifier predictions. My experiments on the natural and adversarial images datasets show that, indeed, there is a visible correlation between the background and the wrong-classified foreground object. This behavior of the network is not supported by human experience, because, for example, we do not confuse the yellow school bus with the snow plow just because it is on the snowy background.
△ Less
Submitted 5 May, 2020;
originally announced May 2020.
