激光与光电子学进展, 2019, 56 (2): 021002, 网络出版: 2019-01-16
基于改进多尺度特征图的目标快速检测与识别算法 下载: 757次
Fast Object Detection and Recognition Algorithm Based on Improved Multi-Scale Feature Maps
图像处理 目标检测与识别 多尺度特征图 卷积神经网络 image processing object detection and recognition multi-scale feature maps convolutional neural network
摘要
针对目标检测与识别在精度和实时性方面的要求, 提出了一种基于改进多尺度特征图的目标快速检测与识别算法。算法在原始SSD模型的基础上, 利用卷积神经网络自动提取多尺度特征图, 构建了一种有效的卷积特征图融合模块, 同时引入轻量级的压缩型双线性融合方法, 丰富上下文信息。进一步结合通道注意机制, 自适应地学习特征图各通道之间的相互关系, 强调有用信息, 抑制冗余信息, 提高了特征图的判别能力, 将增强后的多尺度特征图用于检测模型。实验结果表明, 与同类算法相比, 所提算法的效率更高, 明显提升了识别精度, 同时速度达到63 frame·s-1, 较好地平衡了识别精度与速度之间的关系。
Abstract
Aiming at the precision and real-time requirements of object detection and recognition, a fast object detection and recognition algorithm is proposed based on the improved multi-scale feature maps. In this algorithm, the convolutional neural network is adopted to automatically extract the multi-scale feature maps based on the original SSD model, and a fusion module of effective convolutional feature maps is constructed. Meanwhile, the lightweight compact bilinear fusion method is introduced to enrich the context information. Moreover, through the channel attention mechanism, the relationship among all channels of feature maps is adaptively learned. The useful information is emphasized, in contrast the redundant information is suppressed. Thus the discriminability of feature maps is boosted. The enhanced feature maps are then used in the detection model. The experimental results show that the proposed algorithm is more efficient than the other similar algorithms and the recognition precision is obviously improved, meanwhile, the speed of proposed algorithm is up to 63 frame·s-1. The recognition precision and speed are well balanced.
单倩文, 郑新波, 何小海, 滕奇志, 吴晓红. 基于改进多尺度特征图的目标快速检测与识别算法[J]. 激光与光电子学进展, 2019, 56(2): 021002. Shan Qianwen, Zheng Xinbo, He Xiaohai, Teng Qizhi, Wu Xiaohong. Fast Object Detection and Recognition Algorithm Based on Improved Multi-Scale Feature Maps[J]. Laser & Optoelectronics Progress, 2019, 56(2): 021002.