基于深度特征学习的图像自适应目标识别算法

Image Adaptive Target Recognition Algorithm Based on Deep Feature Learning

针对传统图像识别算法过程繁琐、特征提取困难等问题,提出一种基于深度特征学习的图像自适应目标识别算法。首先对每层网络单个特征图的输入进行批量归一化(BN)处理,其次采用参数化线性修正单元PReLU对参数进行自适应调节,比较BN算法作用在激活函数前后输出的表现性能,构建自适应卷积神经网络模型CNN-BN-PReLU.实验从网络层数、卷积核数目、网络优化及经典卷积神经网络模型4个方面进行比较分析,结果表明,在DDSM数据集上,CNN-BNPReLU较优化前准确率提高了8.5%,训练时间大幅减少71.83%,其敏感度、特异度及AUC值均有显著提升,分别达到了96%,87%和0.91,识别效果远高于LeNet-5和AlexNet,具有较好的应用价值。

Aiming at the complicated process of traditional image recognition algorithm and the difficulty of feature extraction,an image adaptive target recognition algorithm based on deep feature learning is proposed.First,the input of the single feature graph of each layer is batch normalized(BN);then the parametric rectified linear unit(PReLU)is used to adjust the parameters adaptively.The performance of the BN algorithm before and after the activation function is compared and the adaptive convolutional neural network model CNN-BN-PReLU is constructed.The experiments are compared and analyzed from four aspects:the number of network layers,the number of convolution kernels,network optimization and classical convolutional neural network models.The results show that the accuracy of CNN-BN-PReLU is improved by 8.5%,the training time is significantly reduced by 71.83%,the sensitivity,specificity and AUC value of CNN-BN-PReLU are improved by 96%,87%and 0.91,respectively,and the recognition effect is much higher than LeNet-5 and AlexNet,which has good application value.