Survey of Lightweight Object Detection Model Methods

Deep neural networks have recently achieved significant success in many visual recognition tasks. However, these models suffer from being computationally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with strict latency requirements. Therefore, studying and designing lightweight models is one of the current research hotspots. Building a lightweight object detection algorithm is a challenging task because the model must be carefully designed to minimize parameters and computing resources while achieving high detection performance. In the past 10 years, lightweight algorithm design has made significant progress. In this paper, we summarize lightweight object detection algorithms to further explore new ideas and directions for lightweight network design. This paper divides the lightweight algorithm design into three categories: backbone network, feature fusion network, and attention mechanism. For each category, this paper provides a detailed analysis of the performance, advantages, and limitations of the modules in the object detection task. Then, we briefly introduce some recent successful methods, such as the Mobilenet series and DETR. Finally, we conclude this paper and discuss the challenges and potential directions for future work. The development trend of lightweight design in the field of object detection is examined.