一种基于变分辨率机制的机坪目标检测算法

多目标检测任务存在目标尺寸变化范围大的情况。通过采集高分辨率图像，可以保证对小目标的观测效果，但是在原始图像中进行滑动窗口扫描式检索，会造成计算成本的显著增加；大尺度目标的检测通过对原始高分辨率图像压缩快速完成，但压缩过程会导致小目标的大量细粒度特征丢失:因此，文章提出一种基于变分辨率机制的YOLO检测 (multiple resolution mechanism based YOLO，MRMY) 算法,并用深度可分离卷积进行优化。该算法考虑特定场景下不同尺寸目标间的位置关系，采用多分辨率机制，先对高分辨率图像进行压缩，对大目标进行快速检测，再根据大目标位置信息确定小目标的搜索空间，并在原始高分辨率图像的局部区域进行小目标识别。由于在任一分辨率下目标尺寸较为明确，因此可对检测模型基于归一化层剪裁掉网络中不重要通道。尽管检测任务需要2次模型运算才能完成，但通过该检测模型可提高算法的速度。在网络公开数据集和自建数据集上的测试结果表明，MRMY算法的全类平均正确度（mAP）比YOLO V4算法提升约21%，检测速度为84帧/秒与YOLO V4的83帧/秒相近。

Fast Multi-scale Target Detection in Apron Based on Variable Resolution Mechanism

In the apron target detection, the target size is very small and a large number of fine particle characteristics are lost in the compression process, and this results in the recognition errors. Fine granularity directly affects the accuracy of target recognition. However, due to the limited energy consumption and computational power of equipment, and the insufficient feature extraction of small targets by universal target detection algorithms, the improvement of speed and accuracy of such algorithms is restricted. This paper presents a real-time detection algorithm YOLO for apron target detection. By using the variable resolution mechanism, the algorithm detects the compressed image once, and then identifies the difficult targets twice. Finally, the unimportant channels in the network are trimmed based on the scaling factor of the batch normalization layer. The trimmed and slimmed algorithm has a higher speed. The test results on the network public data set and the self-built data set show that the MRMY algorithm improves the mAP by about 21% compared with the YOLO V4 algorithm, and the detection speed is 84 FPS, which is similar to the 83 FPS of YOLO V4.