基于小波变换与卷积神经网络的羊脸识别模型

为解决养殖场条件下羊只的个体识别问题,本文基于小波变换与卷积神经网络,提出一种融合频域特征与空间域特征的羊脸识别模型DWT-GoatNet。首先采集总计30只高相似度西农萨能奶山羊日间、夜间两种不同光照环境下的面部图像,基于SSIM指标剔除其中相似度过高的样本,接着进行图像裁剪,并通过模糊、调整亮度、平移、旋转、加入噪声、缩放等方法完成数据增强;然后,设计基于二维离散小波变换(2D-DWT)与卷积运算的羊脸特征提取模块,完成特征融合;之后,以前述羊脸特征提取模块为基础,添加分类模块,进行卷积神经网络搭建;最后,进行超参数组合寻优,形成羊脸识别模型。试验结果表明,本文所构建的羊脸识别模型在日间、夜间两种不同光照环境下测试集上识别准确率分别可达99.74%和99.89%,高于AlexNet、VGGNet-16、GoogLeNet、ResNet-50、DenseNet-121等经典卷积神经网络模型,说明所构建模型适用于羊只的个体识别,为精准养殖、农险理赔领域相关工作提供了有效解决方案。

Goat Face Recognition Model Based on Wavelet Transform and Convolutional Neural Networks

To recognize an individual goat under farm conditions, a novel goat face recognition model named DWT-GoatNet was proposed based on wavelet transform and convolutional neural networks, which integrated frequency domain features and spatial domain features. Firstly, facial images of a total of 30 highly similar Xinong Saanen dairy goats were collected under two different light conditions, daytime and night. Some images were discarded based on structural similarity (SSIM), and the remaining images were cropped manually. Image sets were also augmented by operations of blur, brightness adjustment, translation, rotation, noise addition and scaling. Secondly, a goat face feature extraction module was designed based on two dimensional discrete wavelet transform (2D-DWT) and convolution operation to achieve feature fusion. Then, with this module, a classification module was added and a convolutional neural network named DWT-GoatNet was built. Finally, the combination of hyper-parameters was optimized and goat face recognition model was formed. The experimental results showed that the accuracy of the proposed goat face recognition model can reach 99.74% and 99.89%, respectively, on test set under different light conditions of daytime and night, which was higher than that of some classical CNNs such as AlexNet, VGGNet-16, GoogLeNet, ResNet-50 and DenseNet-121, while the DWT-GoatNet can provide an effective recognition for some related fields of precision farming and agricultural insurances.