Small Object Detection Based on Bidirectional Feature Fusion and Multi-scale Distillation

In this paper, we present a novel approach to address the challenges of small object detection. The proposed method combines multi-scale distillation and bidirectional feature fusion to enhance the performance of small object detection models. By leveraging the knowledge of a teacher model, the student model captures both high-level and low-level features, leading to improved detection results. Then, bidirectional feature fusion aggregates multi-layer features within the student more effectively via a convolutional attention mechanism and pyramid pooling loss function. Extensive experiments demonstrate that the approach significantly improves small object detection performance over baselines, validating the use of multi-scale knowledge distillation and feature fusion techniques for this task.