Abstract:

To address the problems of catastrophic forgetting， overfitting， and difficulty in adapting dense detector characteristics leading to limited detection accuracy faced by the current time-sensitive target class incremental object detection of UAV images， this paper proposes a time-sensitive target incremental detection algorithm based on multi-stage distillation. The proposed method includes three key modules： the continuous Wasserstein distance-based Inter-Class Distillation （WICD） module， the Prototype-Guided Intra-class Consistency Distillation （PGICD） module， and the Cross-prediction Adaptive Distillation （CAD） module. The WICD module captures the inter-class feature differences from the feature graphs and semantic query vectors using Gaussian distributions with continuous Wasserstein distances to enhance the inter-class discriminability. The PGICD module enhances the inter-class discriminative properties by minimizing the high-level semantic query of the instances in both the teacher’s network and student’s network， and the low-level feature graphs with the prototype differences to achieve effective intra-class feature transfer and enhance intra-class consistency. The CAD module optimizes the cross-prediction distillation process by dynamically adjusting the distillation weights of the classification and regression branches， mitigating the problem of catastrophic forgetting in incremental learning， and improving the model’s detection accuracy in complex scenarios. Experimental results on SIMD and MAR20 datasets show that the proposed method performs well in all types of one-step and multi-step incremental scenarios， and the Average Precision （AP） is significantly improved compared with traditional methods. The AP is as high as 70.8% in the incremental scenario of SIMD dataset with 8+7 classes， with an absolute gap of 1.7% and a relative gap of 2.3% from the upper limit. The results also show that the AP is as high as 60.2% in the incremental scenario of MAR20 dataset with 10+10 classes， with an absolute gap of 2.3% and a relative gap of 3.6% from the upper limit. In addition， the effectiveness of each module is verified by ablation experiments， which effectively improves 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