Feature map quantification: an efficient approach for active trachoma image classification

Convolutional neural networks (CNNs) classify inverted eyelid images related to active trachoma. However, using these complex networks in medical service centers faces challenges due to computational resource constraints. To overcome this challenge, we propose a quantified feature map-based filter pruning framework (FSIM-SVD) that relies on feature similarity and feature map contributions. Based on these insights, our approach involves quantifying redundant feature maps using feature similarity (FSIM) and assessing the contribution of each feature map through singular value decomposition (SVD). By analyzing the impact of each component on the overall model performance, less significant filters can be identified and pruned. The experiment uses VGG16, ResNet56, and ResNet110 on the active trachoma and CIFAR10 datasets. The results reveal that VGG16 achieved an accuracy of 86.9 % (+0.43 % from the baseline) for active trachoma classification while reducing FLOPs by 28.6 % and parameters by 33.4 %. In the CIFAR10 classification, ResNet110 achieved an accuracy of 94.31 % (+0.73 % from the baseline) with a 43.8 % reduction in FLOPs and a 43.1 % reduction in parameters. Compared to state-of-the-art compression techniques, the proposed approach achieves a higher pruning rate and improved classification performance.