Neural network performance effects due to image sampling aliasing error

We explore the effects of image sampling aliasing error on convolutional neural network (CNN) performance is explored, where the aliasing error is quantified utilizing the author’s previous works on sampling and aliasing in digital imaging systems. We start by examining the performance of a single pass, probabilistic neural network (PNN) to understand and appreciate the effect of image sampling aliasing on a simple neural network activation function. This provides direct access to the neural network mechanism for object identification and allows a more direct exploration of the effects of aliasing error on performance. In addition, it is determined that if anti-aliasing filters (low-pass filters) are applied to the input test images, then the neural network performance is significantly improved. This leads us next to a performance evaluation of the pretrained MATLAB CNN with respect to the aliasing error in the input test images. This study’s motivation is to aid in determining sampling requirements for optical imaging systems, in particular that provide input to CNNs from an image aliasing error perspective. We have applied this concept to an optical image (spatial data), but these same concepts can be applied to other imaging data such as radar or LiDAR as well as temporal data such as voice recognition.