基于深度学习的复杂背景下目标检测

Object Detection Based on Deep Learning in Complex Background

针对某些静态图像背景复杂,受环境因素(光照、遮挡、掩盖等)影响较大的问题,提出一种基于深度学习算法的卷积神经网络(convolutional neural networks,CNN)结构对目标进行检测。利用CNN网络可自主提取图像特征并进行学习的优点,避免了复杂的人工特征选择和提取过程。通过一种区域合并的方法进行端到端的交替训练,在复杂背景图像的处理中体现出较优的性能。CNN的局部连接、权值共享及池化操作等特性使之可以有效地降低网络的复杂度、减少训练参数的数目、提高检测效率。试验验证结果表明:此方法在互联网图像数据库检测方面达到了较高的精度。采用坦克模型图像对复杂背景下的单目标、多目标以及不同程度的遮挡、伪装等情况进行试验,得出该方法具有一定的鲁棒性。

Object detection in complex background is one of the key problems in computer vision.The main task is to identify and locate the objects in the image. For some static images with complex background and influenced by environmental factors（light, occlusion, concealment etc.）, a convolutional neural network（CNN） structure based on deep learning algorithm is proposed to detect the object. The CNN network can automatically extract the features of the image and the advantages of learning to avoid the complex artificial feature selection and extraction process; the end-to-end training is carried out by a method of regional merging,which shows better performance in the processing of complex background images. The local connection,weight sharing and pooling operations of CNN can effectively reduce the complexity of the network,reduce the number of training parameters andimprove the detection efficiency. Through the test verification,this method achieves high detection accuracy on the Internet image dataset,and uses the tank model images to test the single target,multi-target and different degree of occlusion and camouflage in the complex background. The method has the advantages of robustness.