Training Neural Networks for Object Recognition Using Blurred Images

A shortcoming of deep neural networks is that they require a vast amount of data to train. An example of this is object recognition, a computer vision technique for identifying objects in images or videos. We hypothesized that, when training with few data examples, blurring the input images would cause the neural network to perform better compared to non-blurred images, because of the removal of unnecessary details. In this study, we trained a convolutional neural network on the blurred images, varying the amounts of blur in order to determine how the validation accuracy changes. Our preliminary results suggest that blurring the images does not help when learning from few examples; however, we cannot fully disprove the hypothesis because it requires further experimentation with other data sets and convolutional neural network models. In the future, we can use image blurring to study eccentricity dependence, a property of the human visual system that standard convolutional neural networks do not currently replicate. This research will ultimately bring us closer to understanding and replicating how the human brain processes vision.