一种基于模型压缩的行人重识别方法

A Person Re-identification Method Based on Compression Model

深度学习模型在行人重识别领域虽已取得较大的发展,但由于对计算力和存储空间的要求比较高,仍然制约着其在开放空间中的应用。针对上述问题,提出一种通道剪枝、分组卷积、自注意力机制优化组合策略进行行人重识别模型压缩,提升模型在开放空间中的精准度和高效率。此方法可根据具体场景的需求灵活组合、获得适用于边缘智能设备的最佳Re-ID模型,具有高效、灵活、易拓展等特性。

Although deep learning model has achieved inspiring progress in person re-identification(Re-ID),its applications in open space remain limited due to high computation and storage.To address this issue,a strategy is proposed to compress Re-ID model by optimally combining channel pruning,grouping convolution and self-attention mechanism,which can immensely improve the accuracy and efficiency of model in open space.The method can be extended flexibly according to the requirements of specific scenarios and obtain the best Re-ID model suitable for edge intelligent devices with the characteristics of high efficiency and flexibility as well as easy expansion.