基于多尺度特征选择与融合的目标检测方法

针对多尺度目标检测中特征图特征混淆和特征丰富程度不足的问题，提出一种基于多尺度特征选择与融合的目标检测算法。设计了一个特征选择模块来分离出不相关的特征，并结合特征金字塔网络形成特征选择网络结构，降低特征图中不同尺度目标的局部特征对当前尺度特征的干扰;提出一种浅层特征融合方法，将浅层特征逐级融合到较深层级特征中，解决特征图的特征不够丰富问题。结合特征选择架构和浅层特征融合架构，在PASCAL-VOC2007数据集上进行测试，结果mAP达到了80.1%。相较于基础的单阶段目标检测(single shot detection,SSD)，所提算法的网络性能可提高2.9%，且在一些小目标和遮挡目标的检测效果上有明显的提升。通过对比和消融实验，证明了所提方法的有效性。

Multi-scale feature selection and fusion for object detection

Aiming at the problem of feature map confusion and lack of feature richness in multi-scale object detection, an algorithm is proposed based on multi-scale feature selection and fusion. First, a feature selection module is designed to separate irrelevant features, and a feature selection network structure is formed by combining the Feature Pyramid Network to reduce the interference of local features of different scale targets in the feature map with current scale features. In addition, a shallow feature fusion method is proposed, which merges shallow features into deeper features step by step, and solves the problem of insufficient feature richness in feature maps. This algorithm is tested on the PASCAL VOC2007 dataset and achieves an mAP of 80.1% by combining the feature selection architecture and the lower feature fusion architecture. Compared with the SSD (Single Shot Detection), the proposed algorithm can improve network performance by 2.9%, and significantly improve the detection performance of some small and occluded targets. The effectiveness of the proposed method is verified by comparison and ablation experiments. The effectiveness of this method is verified through comparison and ablation experiments.