并行多尺度特征增强的小目标检测

由于小目标像素点少，本身携带的特征较少，大多数目标检测算法不能有效利用特征图中小目标的边缘信息和语义信息,导致小目标检测精度低，漏检、误检现象时有发生。为解决RetinaNet模型小目标信息特征不足的缺陷，在RetinaNet模型中引入一个并行辅助的多尺度特征增强模块MFEM(muti-scale feature enhancement model)，通过使用不同扩张率的空洞卷积，避免多次下采样造成信息损失的同时，有利于辅助浅层提取多尺度上下文信息。另外，采用一种专门针对目标检测任务而设计的主干网改进方案，可有效保存高层特征图的小目标信息。传统自上而下的金字塔结构侧重于将高层语义从顶层传递到底层，单向传递的信息流不利于小目标的检测。将辅助MFEM分支与RetinaNet相结合，构造一个包含双向特征金字塔结构的模型，它可有效地融合网络高层强语义信息和底层高分辨率信息。为证明文中算法FE-RetinaNet (Feature Enhancement RetinaNet)的有效性，在MS COCO公共数据集进行实验。与原RetinaNet相比，改进的RetinaNet在MS COCO数据集上检测精度（mAP）取得了1.8%的提升，COCO AP为36.2%；FE-RetinaNet在小目标上检测效果良好，APs提高了3.2%。

Small Target Detection with Parallel Multi-scale Feature Enhancement

Because small targets have fewer pixels and carry fewer features, most target detection algorithms can not effectively use the edge information and semantic information of small targets in the feature map, resulting in low precision of small target detection, and the phenomena of missed detection and false detection occur from time to time. In order to solve the defect of insufficient information features of small targets in RetinaNet model, a parallel assisted multi-scale feature enhancement module MFEM (muti scale feature enhancement model) in RetinaNet model is introduced. By using hole convolution with different expansion rates, it avoids information loss caused by multiple down sampling, and is conducive to assisting in shallow extraction of multi-scale context information. In addition, a backbone improvement scheme specially designed for target detection task is adopted, which can effectively save the small target information of high-level feature map. The traditional top-down pyramid structure focuses on transferring high-level semantics from top to bottom, and the one-way information flow is not conducive to the detection of small targets. The auxiliary MFEM branch with RetinaNet is combined to construct a model containing a bidirectional feature pyramid structure, which can effectively integrate the high-level strong semantic information and the low-level high-resolution information.In order to prove the effectiveness of the proposed algorithm FE-RetinaNet (feature enhancement RetinaNet), experiments are carried out on MS COCO public data set. Compared with the original RetinaNet, the detection accuracy (mAP) of the improved RetinaNet on MS COCO dataset has been improved by 1.8%, and the COCO AP is 36.2%; Fe RetinaNet has a good detection effect on small targets, and APs has increased by 3.2%.