关键词: 增量目标检测, 知识蒸馏, 无人机图像, 时敏目标检测, 交叉预测蒸馏

Abstract: To address the challenges of catastrophic forgetting, overfitting, and limited detection accuracy due to the difficulty in adapt-ing to dense detector characteristics in class-incremental detection of time-sensitive targets in UAV images, this paper pro-poses a time-sensitive target incremental detection algorithm based on multi-stage distillation. The algorithm primarily com-prises a Wasserstein-based Inter-Class Distillation (WICD) module, a Prototype Guided Intra-Class Consistency Distillation (PGICD) module, and a Cross Prediction Adaptive Distillation (CAD) module, with experimental validation conducted on the SIMD and MAR20 datasets. The WICD module captures inter-class feature differences from feature maps and semantic query vectors, leveraging Gaussian distribution and continuous Wasserstein distance to enhance inter-class discriminability. The PGICD module achieves effective intra-class feature transfer and strengthens intra-class consistency by minimizing pro-totype discrepancies between high-level semantic queries and low-level feature maps of instances in teacher and student net-works. The CAD module optimizes the cross-prediction distillation process by dynamically adjusting the distillation weights of classification and regression branches, alleviating catastrophic forgetting in incremental learning and improving the mod-el’s detection accuracy in complex scenarios. Experimental results on the SIMD and MAR20 datasets demonstrate that the proposed method performs excellently in various one-step and multi-step incremental scenarios, with significantly improved average precision (AP) compared to traditional methods. For example, in the 8-class + 7-class incremental scenario on the SIMD dataset, the AP reaches 70.8%, with an absolute gap of 1.7% and a relative gap of 2.3% compared to the upper bound. In the 10-class + 10-class incremental scenario on the MAR20 dataset, the AP reaches 60.2%, with an absolute gap of 2.3% and a relative gap of 3.6%. Additionally, ablation experiments verify the effectiveness of each module, effectively enhancing the incremental detection performance of time-sensitive targets in UAV images.

Key words: Incremental Object Detection, Knowledge Distillation, UAV Images, Time-sensitive Target Detection, Cross Prediction Distillation