The Effect of Changing Image Contrast on Object Recognition by a Convolutional Neural Network

The effect of changing image contrast on the recognition of simple objects by a convolutional neural network is studied using the example of simple high-contrast images. Images of handwritten digits and letters, which have a significantly higher brightness than the surrounding background, are selected as recognizable objects. Images of handwritten digits (MNIST) and Latin letters (EMNIST) measuring 28 by 28 pixels are selected as training and test datasets. Changing the contrast is achieved by reducing the overall brightness of the entire image, as if simulating a decrease in the illumination of the object. It is established that recognition accuracy decreases as the contrast decreases, deteriorating twofold when the contrast changes tenfold. The training of a neural network at various contrast levels is considered. Both increasing and decreasing the contrast are shown to impair recognition. Histograms are presented demonstrating the recognition accuracy depending on the contrast ratio. It is shown that the initial weights of the neural network affect the recognition accuracy.