Back and Forward Incremental Learning Through Knowledge Distillation for Object Detection Unmanned Aerial Vehicles

Incremental object detection is essential for real-world applications where the models must continuously learn new object categories without forgetting previously acquired knowledge. However, the conventional training methods often suffer from catastrophic forgetting, leading to performance degradation on early tasks when learning new ones. To address this, we proposed a novel framework that integrates an attention-based feature enhancement module within a knowledge distillation-based incremental learning pipeline. The AFEM combines self-attention and cross-attention mechanisms to strengthen feature representations by enhancing spatial localization and classification across tasks. Our approach employs a teacher-student architecture, where the teacher model preserves prior knowledge while the student model learns new classes through multi-level distillation losses applied to features, region proposals, and final predictions. Experiments conducted on the Pascal VOC dataset demonstrate the effectiveness of our method. Our method achieves 60.802% AP for the base class. It maintains 56.969% AP after one incremental phase, with a forgetting rate of just 3.833% using Huber loss and achieving improved detection accuracy with MSE loss. These results indicate that our method effectively retains performance on the old classes while adapting to new ones.