Current Issue Cover
一种基于全卷积语义分割网络的物体检测模型“ChinaMM2018”

肖锋,芮挺,任桐炜,王东,孙鹏宇(陆军工程大学;南京大学计算机软件新技术国家重点实验室;中国人民解放军 91292部队)

摘 要
目的 目前主流物体检测算法需要预先划定默认框,通过对默认框的筛选剔除得到物体框。本文提出了一种不需要划定默认框,实现完全端到端深度学习语义分割及物体检测的多任务深度学习模型(FCDN)。方法 首先分析了被检测物体数量不可预知是目前主流物体检测算法需要预先划定默认框的原因;然后提出了一种不需要划定默认框的物体检测算法,该算法采用语义分割思想,在像素级别上先检测出所有的物体边界关键点,再结合语义分割图的类别信息,得到预测框。结果 最后在VOC2007 test数据集上测试,验证了语义分割物体检测方法的可行性,并与目前主流物体检测算法进行了性能对比,结果表明,利用新模型可以同时实现语义分割和物体检测任务,在训练样本相同的条件下训练后,其物体检测精度优于经典的物体检测模型;在算法的运行速度上,相比于FCN,减少了8ms,比较接近于YOLO等快速检测算法。结论 本文提出了一种新的物体检测思路,它充分利用语义分割提取的丰富信息,通过减少语义分割预测的像素点来提高检测效率,并通过实验验证简化后的语义分割结果仍足够进行物体检测任务。
关键词
Object Detection Method Based on Fully Convolution Networks for Semantic Segmentation

Xiao Feng,Rui Ting,Ren Tongwei,Wang Dong,sun peng yu(School of Field works,PLA Army Engineering University,Nanjing;State Key Laboratory for Novel Software Technology,Nanjing University,Nanjing)

Abstract
Objective At present, the mainstream object detection algorithm needs to delimit the default box in advance and get the object box by filtering out the default box. we proposed a multitask learning model (FCDN), which does not need to delimit the default box to realize complete end to end deep learning semantic segmentation and object detection. Method First, the reason that the number of objects being detected is undetermined is the reason that the current mainstream object detection algorithm needs to delineate the default box in advance; then a kind of object detection algorithm which does not need to delimit the default box is proposed, which uses the semantic segmentation idea to detect all the object boundary key points at the pixel level. The ground truth box is obtained by combining the category information of the semantic segmentation result. Result The feasibility of the semantic segmentation object detection method is verified by VOC 2007 test image data sets, and the performance comparison results with the current mainstream object detection algorithm show that the semantic segmentation and object can be realized at the same time by using the new model, trained with the same training sample, detection precision of FCDN is superior to that of classic detection models. In terms of the running speed of the algorithm, compared with FCN, it reduces by 8ms, which is close to the fast detection algorithm such as YOLO. Conclusion This paper puts forward a new thought of object detection, it take full advantage of the semantic segmentation to extract information, by reducing the pixels of semantic segmentation prediction to improve the efficiency of detection, and through the experiment of simplified semantic segmentation result is still enough for object detection task.
Keywords
QQ在线


订阅号|日报