Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{203952,
        author = {K.Mounika and S.Sarojini Devi},
        title = {PaPr: A Study on Critical Architectural Design and Optimization for Training-Free Patch Pruning via Entropy-Based Dynamic Control and Hybrid Saliency Integration},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {1},
        pages = {2013-2020},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=203952},
        abstract = {The Rapid advancements in Deep Neural Network (DNN) architectures have led to significant computational challenges due to the need for precise inference. Methods such as Training-Free Patch Pruning (PaPr) have been employed to address these challenges without requiring costly retraining, but they still suffer from two critical drawbacks: Saliency Drift and Budget Rigidity. This paper proposes an architecture-based solution for overcoming these limitations. Entropy-Aware Dynamic Controller is presented to incorporate information theory metrics in adapting the patch retention budget based on scene entropy to replace rigid keep ratios. On the other hand, Hybrid Saliency Fusion method is adopted to alleviate the problem of saliency drift by utilizing feature maps produced by a lightweight scorer and decision gradients from backbones extracted via Grad-CAM. The results from experiments on the ImageNet-10k dataset using a MobileNetV2-ResNet50 dual-model framework demonstrate that our proposed hybrid approach achieves about 98% recovery rate in scenes with high entropy levels.},
        keywords = {Patch Pruning, Model Compression, Saliency Drift, Entropy-Aware Control, Hybrid Saliency Fusion, Grad-CAM, Deep Learning Inference, Edge Computing.},
        month = {June},
        }