Model pruning method for differentiated image classification tasks on edge devices

Yanting Wang

doi:10.1016/j.sysarc.2026.103782

Return

Model pruning method for differentiated image classification tasks on edge devices

2026-03-16

PRE

DOI：10.1016/j.sysarc.2026.103782

Original

Original request for help

Save

Abstract

En 中文

Deploying large-scale models on resource-constrained edge devices is a challenge, particularly when models trained for general-purpose classification must be specialized for tasks involving only a subset of classes. Existing pruning methods are predominantly task-agnostic and fail to account for task-specific class composition, resulting in ”task-level category redundancy” where irrelevant model components are retained and thus limit achievable compression. To address this, we propose Task-Specific Pruning (TSP), a method that integrates task-category information into pruning pipeline. Specifically, TSP quantifies importance of each filter with respect to individual classes using a rank-based feature map analysis, and derives task-specific importance scores by fusing class-level information according to target task. To improve stability under aggressive compression, a layer-wise minimum retention constraint is further introduced. Extensive experiments on CIFAR-10, CIFAR=100, and ImageNet-1K, across multiple backbone architectures including VGG, ResNet, and MobileNet, demonstrate that TSP consistently outperforms existing methods in pruning rate and FLOPs reduction under accuracy-loss constraints of 0.5%, 3%, and 10%. Deployment on NVIDIA Jetson Orin Nano further demonstrates significant inference latency reductions on real edge hardware, underscoring the necessity of system-level evaluation beyond FLOPs. These results highlight the importance of explicitly modeling task-category characteristics for efficient and reliable model specialization in edge scenarios.

Keywords:

Task-Specific Pruning

Edge Device Deployment

Model Compression

Class-Selective Pruning

FLOPs Reduction