This is the source code for our paper: 基于知识蒸馏的目标检测模型增量深度学习方法.

随着万物互联时代的到来，具备目标检测能力的物联网设备数量呈爆炸式增长。基于此，网络边缘产生了海量的实时数据，具有低时延、低带宽成本和高安全性特点的边缘计算随之成为一种新兴的计算模式。传统的深度学习方法通常假定在模型训练前所有数据已完全具备，然而实际的边缘计算场景中大量的新数据及类别往往随时间逐渐产生和获得。为了在训练数据成批积累和更新的条件下在资源有限的边缘设备上高效率地完成目标检测任务，本文提出了基于多中间层知识蒸馏的增量学习方法（incremental learning method based on knowledge distillation of multiple intermediate layers,ILMIL）。首先，为了能够适当地保留原有数据中的知识，提出了包含多个网络中间层知识的蒸馏指标（multi-layerfeaturemapRPNandRCN knowledge,MFRRK）。ILMIL将教师模型和学生模型的中间层特征的差异加入模型训练，相比于现有的基于知识蒸馏方法的增量学习，采用ILMIL方法训练的学生模型可以从教师模型的中间层学习到更多的旧类信息来缓解遗忘；其次，ILMIL利用MFRRK蒸馏知识完成现有模型的增量训练，避免训练使用多个独立模型带来的资源开销；为进一步降低模型复杂度以高效地在边缘设备上部署推理，可在知识蒸馏前进行剪枝操作来压缩现有模型。通过在不同场景和条件下的实验对比，本文方法可在有效降低模型计算和存储开销的前提下，缓解已有知识的灾难性遗忘现象，并维持可接受的推理精度。

With the advent of the Internet of Everything era, the number of IoT devices with object detection capabilities has grown exponentially. Consequently, massive amounts of real-time data are generated at the network edge, making edge computing—an emerging computing paradigm characterized by low latency, low bandwidth costs, and high security—increasingly relevant. Traditional deep learning methods typically assume that all data is fully available before model training. However, in real-world edge computing scenarios, large amounts of new data and categories are often generated and acquired gradually over time. To efficiently perform object detection tasks on resource-constrained edge devices under conditions where training data is accumulated and updated in batches, this paper proposes an incremental learning method based on knowledge distillation of multiple intermediate layers (ILMIL). First, to appropriately retain knowledge from existing data, a distillation metric incorporating knowledge from multiple intermediate network layers (multi-layer feature map RPN and RCN knowledge, MFRRK) is introduced. ILMIL incorporates the differences in intermediate layer features between the teacher and student models into the training process. Compared to existing incremental learning methods based on knowledge distillation, the student model trained with ILMIL can learn more information about old classes from the teacher model’s intermediate layers, mitigating catastrophic forgetting. Second, ILMIL leverages MFRRK distillation to perform incremental training of the existing model, avoiding the resource overhead of training multiple independent models. To further reduce model complexity for efficient inference deployment on edge devices, pruning operations can be applied to compress the existing model before knowledge distillation. Through experimental comparisons across different scenarios and conditions, the proposed method effectively reduces computational and storage overhead while mitigating catastrophic forgetting of existing knowledge and maintaining acceptable inference accuracy.

本文发表在工程科学与技术，链接，被评为该刊2024年高影响力论文(证书)。

@article{方维维2022基于知识蒸馏的目标检测模型增量深度学习方法,
title={基于知识蒸馏的目标检测模型增量深度学习方法},
author={方维维 and 陈爱方 and 孟娜 and 程虎威 and 王清立},
journal={工程科学与技术},
volume={54},
number={6},
pages={59--66},
year={2022},
publisher={工程科学与技术}
}

Please note that the open source code in this repository was mainly completed by the graduate student author during his master's degree study. Since the author did not continue to engage in scientific research work after graduation, it is difficult to continue to maintain and update these codes. We sincerely apologize that these codes are for reference only.