Abstract:

In order to improve image object recognition speed for multi-attributes of multi-values, an end-to-end recognition algorithm is proposed. Firstly, the modified YoloV3 network is used as main network in order to detect the object bounding boxes. Sub-networks are constructed according to the independent attributes. Sub-networks share the deep bounding box features of main network and adopt multi-outputs to recognize the attributes multi-values. There are three stages with different objective functions in the training process. Experimental results show that the proposed algorithm has good performance.

Key words: object detection, attribute recognition, deep learning, convolutional neural network, image recognition

摘要：

为提高图像目标多种多值属性的识别速度，提出一种端到端的识别算法。采用修正的YoloV3网络作为主网络，确定目标的boundingbox；依据属性独立特性构造子网络，多个子网络共享由boundingbox确定的主网络深层次特征，进行推断，并采用多值输出满足多值属性的识别。在训练过程中，采用了三阶段分目标训练。实验结果验证了该算法在识别准确度和时间效率上的优良性能。

关键词: 目标检测, 属性识别, 深度学习, 卷积神经网络, 图像识别