An Object Detection Algorithm Based on Attention Mechanism and Lightweight Network (AMLN)

Abstract

In recent years, with the widely accepted and in-depth research of deep network and deep learning, the techniques of object detection have been greatly promoted. Visual attention mechanism is an important visual signal processing approach of human beings. The incorporated of attention mechanism into computer vision applications based on deep learning, for various tasks such as image classification, object detection, image segmentation and motion recognition, has reached remarkable achievement with excellent effects. In current, object detection algorithms based on deep learning can be implemented by two main methods: candidate region method (two-stage method) and regression method (one-stage method). The former method -- two-stage object detector can acquired more detection accuracy than the later method -- one-stage object detector, however with inherent drawback -- slower speed to derive detection conclusion. This paper proposes a new object detection algorithm (AMLN) which incorporating the Attention Mechanism into the two-stage detection model, replacing the heavy design architecture for detection component with Lightweight Networks, and adopting soft-NMS instead of NMS as post-processing for prediction component. The experiments are evaluated on the COCO dataset, and the results show that the proposed object detection algorithm based on Attention Mechanism and Lightweight Network (AMLN) can achieve significantly higher detection efficiency while maintaining higher accuracy.