Convolutional Neural Network Features Selection Based On Analysis Of Variance
Bachir Kaddar, Hadria Fizazi, Dou El Kefel Mansouri · 2019
In this paper, we aim to reduce the redundancy in deep Convolutional Neural Networks (CNNs) while preserving significantly the image classification performance. The main novelty is to select a subset of most discriminative features in the final response layer by pruning neurons with the least importance. To achieve that, the Analysis of Variance (ANOVA) technique is used to determine feature maps with higher neuron responses variation. The proposed method handles effectively the neurons with the least important and may prove useful in identifying the image regions most informative regarding classification. Experimental results, using several datasets, show that our proposed neurons pruning strategy yields a significant improvement in feature selection ability, and then fine-tuned network to retain its predictive power.