An Improved Convolution Neural Network for Object Detection Using YOLOv2
Enzeng Dong, Yanfang Zhu, Yuehui Ji, Shengzhi Du · 2018
You Only Look Once (YOLO) is an object detection system with real-time effects. It is difficult to accurately detect small targets in complex scenes, so an improved network structure of YOLOv2 is proposed in this paper. The proposed network is improved in the following aspects: firstly, a 1×1 convolutional layer is added to improve the detection accuracy; secondly, the output sizes of several layers are changed from 13×13 to 26×26 to extract more features from multi-pixels image; thirdly, the loss function is also optimized to adapt to the size of objects in the image. Finally, the model is trained and tested based on VOC2007 dataset, and the experiment results show that the improved model has higher accuracy on detecting small objects.