AF-RetinaNet:一种基于自适应融合与特征细化的微小行人检测算法

为解决目前目标检测算法在微小行人的识别与定位过程中准确率较低的问题,提高微小行人检测能力,提出一种基于自适应融合与特征细化的微小行人检测算法AF-RetinaNet.首先,将特征增强模块与ResNet相结合构建特征提取网络,采用并行结构获得增强特征;其次,使用上下文自适应学习模块,通过获得目标上下文的特征信息,从而关注相似特征的差异性,缓解误检问题;最后,构造具有图像超分思想的特征细化模块,对目标特征信息进行放大重构,优化小目标的特征表达能力,缓解漏检问题.在TinyPerson数据集上,AF-RetinaNet算法的检测精度达到56.78%,漏检率达到85.38%.与基于RetinaNet算法的研究基准相比,检测精度提高5.57%,漏检率降低3.67%.实验结果表明,该模型能有效提高对微小行人的检测和识别精度.

AF-RetinaNet: A tiny person detection algorithm based on adaptive fusion and feature refinement

In order to solve the problem of low accuracy of current target detection algorithms in the process of tiny person recognition and location, and improve the detection ability of tiny person, this paper proposes a tiny person detection algorithm AF-RetinaNet based on adaptive fusion and feature refinement. Firstly, the algorithm combines the feature enhancement module with ResNet to build a feature extraction network and uses a parallel structure to obtain enhanced features. Secondly, the context adaptive learning module is used to obtain the feature information of the target context, so as to pay attention to the differences of similar features and alleviate the problem of false detection. Finally, the feature refinement module with the idea of image super-resolution is constructed to enlarge and reconstruct the target feature information, optimize the feature expression ability of small targets and alleviate the problem of missed detection. On the TinyPerson dataset, the average precision of the AF-RetinaNet algorithm reaches 56.78%, and the missed rate reaches 85.38%. Compared with the research benchmark based on the RetinaNet algorithm, the average precision is improved by 5.57%, and the missed rate is reduced by 3.67%. The experimental results show that the model can effectively improve the accuracy of tiny person detection and recognition.