一种融合小波变换与卷积神经网络的高相似度图像识别与分类算法

针对特定领域高相似度图像识别与分类问题，提出融合小波变换与卷积神经网络的高相似度图像识别与分类算法。首先，利用小波变换提取图像纹理特征，对不同类别、不同分辨率图像集进行训练并确定最佳纹理差异度参数值；其次，根据纹理差异度运用小波分解方法对图像进行子图分解，提取各子图能量特征并进行归一化处理；接着，通过卷积神经网络5层卷积和3层池化交替，将输入图像特征向量转化为一维向量；最后，通过训练次数的增加以及数据量的增大，不断优化网络参数，提高在训练集中的分类准确度，在测试集中验证权值实际准确度，得到具有最高分类准确率的卷积神经网络模型。实验选取鸡蛋、苹果两类图像数据集作为实验数据，进行鸡蛋散养或圈养识别、苹果产地判定，实验结果表明：该算法平均鉴别准确率均达90%以上。

A high similar image recognition and classification algorithm fusing wavelet transform and convolution neural network

A high similar image recognition and classification algorithm fusing wavelet transform and convolution neural network is proposed for high similar image recognition and classification in specific fields with small color and texture feature differences. Firstly, image texture featuresare extracted by wavelet transform, and the optimal texture difference parameter threshold is determined by different categories and different resolution image sets. Secondly, the wavelet decomposition method is used to segment the image,extract each subgraph’s energy features, and normalize them. Then, a convolution neural network with 5 convolution layers and 3 pool layers are used to transform the input image texture feature vector into one dimensional vector. Finally, by increasing the training number and the data amount, the network parameters are continuously optimized and the classification accuracy in the training set is improved. The actual accuracy of the weights is verified in the test set, and the convolutional neural network model with the highest classification accuracy is obtained.Eggs and apples are chosen as the experimental data. Whether the eggs are free range or captive and where are the original places of the apples are identified in the experiments. The experimental results show that the average accuracy rate of the algorithm is above 90%.