RefineDet网络与注意力机制结合的目标检测算法

Target detection algorithm combining RefineDet network and attention mechanism

针对单框架算法在目标检测中对小目标检测精度较差的缺点,以Refine Det网络为基础,提出了一种二次定位的检测模型。在融合各层特征的同时加入了注意力机制,从而增强对目标的检测能力。框架通过由粗到细的定位方法,在速度上保持了单框架的优势并且在检测精度上得到了明显提高,对遮挡和小型目标较多的图像依然能获得高品质的检测框。

Aiming at the shortcomings of low detecting precision of small target single-frame algorithm in target detecting,based on RefineDet network,a quadratic positioning detection model is proposed,which fuses the attention mechanism while enhancing the characteristics of each layer,so as to enhance ability to detect targets.The frame maintains the advantages of single frame in speed by coarse-to-fine positioning method and has a significant improvement in detection precision,and still obtains high-quality detection frame for images with more occlusion and small targets.