A Lightweight Adaptive Pruning Model Based on Scene Complexity
BAO Juntao
WU Yue
LIU Bingyou
Abstract：To meet the requirements of efficient deployment of deep learning models in resource-constrained environments,the adaptive scene complexity-based mobile vision transformer(AC-MViT)pruning model was proposed.Firstly,a scene complexity evaluation module was introduced into the mobile vision transformer(MobileViT)network to dynamically adjust the number of channels and transformer layers according to the image complexity,so as to achieve the effect of reducing computational costs in simple scenes and retaining feature details in complex scenes.Meanwhile,the coupling strategy of dynamic convolution and transformer layers was applied,and random pruning and computational cost penalty were added during training to enhance the robustness of the model.The results showed that AC-MViT model had achieved remarkable results on the common objects in context 2017(COCO 2017)dataset and the pattern analysis,statical modeling and computational learning visual object classes 2012(PASCAL VOC 2012)dataset.Compared with MobileViT model,the mean average precision(mAP)decreased by 1.1%and 0.1%on COCO 2017 and PASCAL VOC 2012 dataset respectively,the floating-point operation speed was reduced by 43.1%and 46.6%respectively,the number of parameters decreased by 49.1%and 50.0%respectively,and the inference time was shortened by 48.4%and 52.5%respectively.The AC-MViT model performed excellently under various scene complexities,demonstrating its efficiency and balance in resource-constrained environments applications.
Keywords：deep learningobject detectionadaptive pruninglightweightMobileViTcomplexity evaluation
Publication Date：2025-03-19
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：6( 47-52 )