文献类型：期刊文献

英文题名：Progressive Query Pruning for Object Detection

作者：Qian, Zehong[1]; Cheng, Hua[1]; Mao, Shucheng[1]; Ding, Yingying[1]

机构：[1] School of Information Science and Engineering, East China University of Science and Technology, Shanghai, China

年份：2025

起止页码：400

外文期刊名：2025 6th International Symposium on Computer Engineering and Intelligent Communications, ISCEIC 2025

收录：EI(收录号：20261520488616)

语种：英文

外文关键词：Error detection - Object recognition - Query languages - Query processing - Structured Query Language

摘要：The DETR and its variants achieve end-to-end object detection, but the fixed number of object queries substantially exceeds the actual demand, resulting in redundant or ineffective query focus. To address this issue, this paper proposes a Progressive Query Pruning (PQP) method. Specifically, an adaptive query quality estimator is integrated after each layer of the Transformer decoder to dynamically assess query importance. The estimator jointly evaluates classification confidence, attention concentration, and query distinctiveness to compute an importance score. Based on this score, PQP progressively prunes low-potential and highly redundant queries across layers. Experimental results demonstrate that the proposed method introduces no additional parameters, ensures high-quality query retention, and enhances both detection accuracy and model robustness. ? 2025 IEEE.