Object Detection Based on Deep Learning of Small Samples

Object detection of indoor scene is widely used in the field of service robot. State-of-art object detectors rely heavily on large-scale datasets like PASCAL VOC2007, VOC2012. However, these approaches fail to indoor scene object detection limited by a few samples and the complex background. This paper presents an object detector based on deep learning of small samples. Firstly, the algorithm can augment training samples automatically by synthetic samples generator to solve the problem of few samples. Synthetic samples generator is designed by switching the object regions in different scenes. Then, deep supervision learning and dense prediction structure are used in the deep convolution neural networks. It is a better solution to solve the vanishing-gradient and the objects with different scale. In addition, the semantic relevance of objects is used to improve the accuracy of weak feature objects in complex scenarios. Experiments on B3DO demonstrate that the proposed algorithm achieves better results than the state-of-art contrast models, and the mean average precision (mAP) was 0.18 higher than the DSOD.