Resnet-based quality evaluation of image mosaic and training data balance

Many image Mosaic algorithms have been pro-posed in recent decades, we also need the algorithm to evaluate the quality of image stitching. This paper verifies the possibility of image mosaic quality evaluation based on neural network. Experiment results show it is effective and has the strong scalability on different data sets. Furthermore, the influence of training data balance and the network deep of ResNet are analyzed. Ten times amount data will reduce the mean error from 2 to 1. Balance data set can reduce the mean error significantly. The deeper ResNet34 is better than ResNet18 which can train faster but needs more memory space.