一种面向工业部署的目标检测模型蒸馏技术

Object Detection Models Distillation Technique for Industrial Deployment

深度学习目标检测模型的应用场景相当广泛,然而,受制于部署设备的性能,部署模型的检测精度往往较低。为提高检测模型的性能,本文提出一种高效的动态蒸馏训练方法。该方法创新性地引入动态样本分配策略来筛选教师模型的高质量输出,并配合蒸馏损失的动态权重调整,对传统的目标检测模型蒸馏算法进行改进。在电网安全施工场景数据集上的实验结果表明,相较于直接训练,该方法使YOLOv6-n模型的AP (Average Precision)值平均提高了2.63个百分点。本文提出的蒸馏方法不影响原有部署模型的推理速度,有助于提升目标检测模型在各种工业场景上的检测性能。

The application scenarios of deep learning object detection models are quite extensive.However,the detection accuracy of deployed models is often low due to the performance limitations of deployment devices.To enhance the performance of detection models,this paper proposes an efficient dynamic distillation training method.This method innovatively introduces a dynamic sample assignment strategy to select high-quality outputs of the teacher model,and pairs this with dynamic weight adjustment of distillation loss,thereby improving the traditional distillation algorithm used in object detection models.Experimental results on a dataset for electrical grid safety construction indicate that,compared to direct training,this method increased the Average Precision(AP)value of the YOLOv6-n model by an average of 2.63 percentage points.The distillation method proposed in this paper does not affect the inference speed of the original deployment model and helps to enhance the detection performance of object detection models in various industrial scenarios.