A Lightweight Detection Method for Power Line Fittings Based on Knowledge Distillation
摘要
电力线路金具的实时检测对保障电网安全运行至关重要,但在复杂自然环境下同时满足高精度与高实时性仍颇具挑战:深度检测模型往往参数庞大、计算开销高,而金具种类繁多、场景复杂.为此,本文提出一种基于知识蒸馏的金具部件检测轻量化方法.该方法设计轻量综合知识传输模块(lightweight integrated knowledge transfer module,LIKTM)与任务自适应区域感知蒸馏(task-adaptive region-aware distillation,TARAD),在保持精度的前提下显著降低模型规模并提升推理速度.LIKTM由轻量共享知识编码器(lightweight encoder for shared knowledge,LESK)与平均原型传输策略(lightweight averaged prototype transfer,LAPT)组成,可压缩特征冗余、提炼通用知识并缓解蒸馏信息损失.TARAD针对分类头与回归头设计区域自适应策略,选择性传递教师模型在关键候选区域的软标签与定位提示.实验结果表明,所提方法在多组教师-学生组合下均取得了最高的平均精度均值(mean average precision,MAP),相较基线显著提升检测精度;同时模型参数量平均压缩至原来的约 80%,推理速度最高提升近 4 倍,实现了高精度与高效轻量化的兼顾.
The real-time detection of power line fittings is very important to ensure the safe operation of the power grid.However,achieving both high precision and real-time performance in complex natural environments remains challenging.Depth detection models often have large parameters and high computational overhead,while the fittings have a wide variety and complex scenarios.Therefore,this paper proposes a lightweight detection method for hardware components based on knowledge distillation.The method introduces a lightweight integrated knowledge transfer module(LIKTM)and task adaptive region aware distillation(TARAD),which significantly reduce model size and improve inference speed without sacrificing accuracy.LIKTM is composed of a lightweight encoder for shared knowledge(LESK)and a lightweight averaged prototype transfer(LAPT)strategy,which can compress feature redundancy,refine general knowledge,and alleviate the loss of distillation information.TARAD designed a regional adaptive strategy for the classification head and regression head,and selectively transferring soft labels and localizaiton cues of the teacher model in key candidate regions.Experimental results show that the proposed method achieves the highest mean average precision(MAP)under multiple groups of teacher-student combinations,which significantly improves the detection accuracy compared with the baseline.At the same time,the average number of model parameters is compressed to about 80%,and the inference speed is increased by nearly 4 times,achieving both high precision and high efficiency and lightweight.