Parallel Connecting Deep and Shallow CNNs for Simultaneous Detection of Big and Small Objects
In order to improve the real-time and accuracy of Faster RCNN (Region based Convolutional Neural Networks) for detecting small object, a novel object detection model is proposed in this paper. Our model not only keeps the detection accuracy for big object, but also improves significantly the accuracy for small object, and with very little reduction in term of detection speed. Firstly, a shallow CNN is designed and connected with an improved deep CNN by using skip-layers connection method, which makes full use of the convolution characteristics with different layers to improve the detection ability for small object; Secondly, the detection accuracy of our model is improved further by incorporating the region proposal mechanism in Faster R-CNN, and using 12 kinds of anchors to generate object candidates; Finally, a dimensional reducer is designed by connecting ROI-Pool layer and 1×1 convolutional layer, which accelerates the detection of overall network. The test results on image datasets PASCAL VOC and MS COCO show that the detection accuracy of our model is higher than some current advanced models, and small objects is significantly improved.
Object detection Convolutional neural networks Region proposal Skip-layers connection
Canlong Zhang Dongcheng He Zhixin Li Zhiwen Wang
Guangxi Key Lab of Multi-source Information Mining and Security,Guangxi Normal University,Guilin,Chi College of Computer Science and Communication Engineering,Guangxi University of Science and Technolo
国际会议
广州
英文
78-89
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）