关键词: 深度学习, 卷积神经网络, 航空器检测, 目标检测识别, 图像分类

Abstract: An end-to-end framework based on deep convolutional neural networks which focuses on implementing both detection and recognition tasks of some specific object on large-scale pictures was proposed. Firstly, it frames object detection as a regression problem to spatially separated bounding boxes, and predicts bounding boxes based on a 24-layer convolutional neural network. After solving the regression problem of bounding boxes, a sub-framework was presented which could be implemented with several typical classification networks. Such framework was applied to efficiently detect and recognize aircraft in large-scale images which might be token by remote sensing satellites. Since taking full advantage of the powerful GPU, experiments on private dataset which meets the specific application scenario showed that the efficiency of detection net speed up a lot compared to classical algorithms.

Key words: deep learning, convolutional neural networks, aircraft detection, object detection and recognition, image classification