Deep learning algorithms have made remarkable progress on the object detection based on huge amount of images. However, it is really difficult to train a model with well generalization for the small-scale images with the limited computation resources. To address the problem, BackgroundNet is proposed to guide the learning process of deep learning-based object detection by the extracted information from the background images. The corresponding network learns not only the features of objects, but also the difference between the object and the non-object area, with the purpose of improve the classification performances for small-scale datasets. Based on YOLO, the background images are employed as the extra input data, and then the input layer of BackgroundNet is not traditionally three RGB but six channels. The experimental results done for coal mine dataset and six public datasets show that the proposed method has better performance when dealing with the object detection with small-scale images and their AP-values are averagely larger than YOLO about 27.8% for six public datasets.
展开▼
