Homography transform enhance CNN prediction accuracy on image classification
Chang Li · Applied and Computational Engineering · 2023
Convolutional Neural Network (CNN) image classification is a well-established algorithm that has been implemented in many fields. Benefit from the digitalization process and the exponential increase in the base of smart devices, this algorithm can be applied to even more traditional or casual contexts in the future driven by the trend of Internet of Things. Thus, the needs for optimizing image classification in specific domains may have turned out to be ongoing valuable research direction. This paper focuses on providing optimization under one example context which is using traffic signs as the experimental target. For the randomly selected traffic sign samples in the experiment, the accuracy obtained from the samples treated by the homography transform compared to control group passed all three statistical tests: Two Sample t-test, McNemar's test, and Fisher's exact test. Therefore, research direction has achieved small-scale validation and presents optimism for large-sample experiments and further research in optimization using the introduced strategy in the paper.