Performance of Pre-Learned Convolution Neural Networks Applied to Recognition of Overlapping Digits
Kazuyuki Hara, Daigo Shii, Ryosuke Miyoshi · 2020
We analyzed the performance of pre-learned convolution neural networks (CNN) learned with a single-digit image dataset when they were used to recognize images containing two overlapping digits. The pre-learned network was learned using the MNIST database, and the network architecture was the LeCun network. The overlapping digit images were made using images from the MNIST database. Our goal was to clarify two issues: (1) can a network learned for recognition of single digits in an image classify an image that includes two overlapping digits without doing additional learning using a dataset composed of two overlapping digit images? (2) Is a convolutional neural network (CNN) capable of processing stereoscopic vision? If (1) is possible with a CNN, then we don't need to train a huge number of images that include various combinations of more than two digits. If (2) is possible, stereoscopic vision is also possible without having to learn overlapping images. Our results support the conclusion that stereoscopic vision or a similar function is involved in the CNN.