CNN Descriptor Improvement Based on L2-Normalization and Feature Pooling for Patch Classification

The L2-normalization and feature pooling have a wide range of applications in image classification and have also achieved remarkable results. However, there is much room for the existing descriptors which are extracted from pre-trained Convolutional Neural Network (CNN) models, to meet the requirement of precision for patch classification. We generate CNN descriptors by using L2-normalization and feature pooling on the existing pre-trained CNN descriptors. By evaluation on the Brown dataset, the mean Average Precision (mAP) of descriptor, which is based on Inception-v3 model that applies both L2-normalization and feature pooling, reaches 99.27%, 98.97% and 98.02% in three sub-datasets. Compared with the pre-trained CNN descriptors without L2-normalization and feature pooling, the mAP of pre-trained CNN descriptors can be respectively increased by $1.41 \% \sim 17.31 \%, 3.11 \% \sim 17.99 \%$. $1.19\%\sim15.24\%$. According to the experimental results, it is obvious that L2-normalization and feature pooling are beneficial to improve the performance of pre-trained CNN descriptors.