Verification of Normalization Effects Through Comparison of CNN Models
Kang-Bae Lee, Sang Ha Sung, Doohwan Kim, Sungho Park · 2019
Recently, there has been an increase in the interest in and necessity of image data analysis. Many studies have proposed algorithms for analyzing image data. Accordingly, image data algorithms have kept advancing as a result of new structures or data processing techniques. Ultimately, the level of image processing technology has exceeded human recognition. Nevertheless, neural networks cannot be perfectly applied to image analysis in every field. Many attempts are being made to overcome this limitation. Normalization is one of those attempts. Normalization is one of the methods used for improving the network performance, and it is adopted for several models. The performance of local response normalization is still under dispute. This type of normalization has been insufficiently studied. This study verifies the real effects of local response normalization by using various datasets and conducts an experiment to identify its impact on learning time. The result of this study is expected to demonstrate the effect of local response normalization on image classification.