A Lightweight Single Stage Small Object Detection Network Based on Knowledge Distillation and Novel Feature Networks

Abstract:

With the increasing demand for traffic sign detection, the challenge of small target detection has become particularly prominent. The present study proposes an innovative approach by integrating knowledge distillation, L2 loss function, and CBAM attention mechanism to effectively tackle this issue. With the help of knowledge distillation technology, the knowledge body of a large and complex teacher model can be refined and simplified into a student model, which effectively improves the prediction accuracy of the model and greatly facilitates the development of lightweight models. At the same time, the L2 loss function is introduced. The feature learning ability of the model is strengthened effectively, making it more adaptable and robust. The application of CBAM attention mechanism enhances the model's perception of key targets and enables the model to focus more on key regional features. This series of improvements not only provides a new idea for small target detection but also brings significant performance improvement in actual traffic scenes. Then, the integration method of bidirectional feature pyramid network (BiFPN) is used to enhance the flexibility of neural network to deal with input of different scales, while speeding up and improving the process of feature fusion. The experimental results demonstrate that when processing CCTSDB open traffic sign data set and executing FLOW-IMG small target detection task, the optimized algorithm shows obvious performance improvement, and its accurate recognition rate jumps to 97% and 84.9% respectively For the basic algorithm, two datasets achieved improved accuracy by an innovative approach, improving accuracy by 5.8% and 1.3%, respectively. In terms of resource efficiency, compared to the original teacher model, the newly constructed model reduced the number of computing participants by approximately 15% during execution, while successfully reducing the overall computing task load by 14%.

Key words:YOLOV5, Small target detection, knowledge distillation, CBAM, L2 loss, BiFPN