基于ReLU函数的卷积神经网络的花卉识别算法  

A Recognition Algorithm of Flower Based on Convolution Neural Network with ReLU Function

作　　者：郭子琰 舒心 刘常燕 李雷[1] GUO Zi-yan;SHU Xin;LIU Chang-yah;LI Lei(Nanjing University of Posts and Telecommunications ,Nanjing 210023 ,China)

机构地区：[1]南京邮电大学,江苏南京210023

出　　处：《计算机技术与发展》2018年第5期154-157,163,共5页Computer Technology and Development

基　　金：国家自然科学基金(61070234;61071167;61373137);国家大学生创新创业训练计划项目(SZDG2016024)

摘　　要：目前对于花卉识别的工作较少,且在已有实验结果中,识别准确率和识别计算速度偏低,需要改进算法、改善实验结果。卷积神经网络由于其可以将图像直接作为输入对象从而避免人工提取特征过程的误差,且在各种外物因素下(光照、旋转、遮挡等)具有良好的鲁棒性,所以在图像识别方面具有巨大的优势。因此选取卷积神经网络对花卉进行识别。在传统卷积神经网络中,一般选用Sigmoid函数作为激活函数,但是使用这种函数需要进行预训练,否则将会出现梯度消失无法收敛的问题。而采用近似生物神经激活函数ReLU则可以避免这一问题,提高机器学习的效果和速度。最终达到了92.5%的识别正确率。There is less work on flower recognition presently, and in the existing experiments, the recognition accuracy and recognition calculation speed are low, so it is needed to improve the algorithm and experimental results. Convolution neural network has robustness in various external factors （ illumination, rotation, occlusion, etc. ） and great advantages in image recognition, which can be selected for rec- ognition of flower. In traditional convolution neural network, the Sigmoid function is normally used as the activation function, but it needs to be pre-trained,otherwise there will exist the problem of gradient vanishing and not converging. ReLU function,which is a kind of ap- proximate biological nerve activation functiun,is applied to improve the effect and speed of machine learning and achieves the 92.5% recognition accuracy finally.

关 键 词：ReLU函数 卷积神经网络 花卉识别 近似生物神经激活函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]