液晶与显示, 2020, 35 (4): 389, 网络出版: 2020-05-30
基于Inception V3的图像状态分类技术
Image classification technology based on inception V3
特征提取 图像识别 卷积神经网络 模型训练 feature extraction image recognition convolutional neural network model training
摘要
为了实现对物体状态的分类识别, 本文在GoogLeNet的Inception V3模块基础上进行了优化, 使用Tanh作为激活函数并结合RMSprop, SGD优化器提升了模型的准确率。首先采用三次卷积插值, GAN对图像集进行预处理, 再利用Inception对图像进行训练, 最后结合RMSprop和SGD优化器对模型进行优化。用本文提出的模型在20个烹饪对象的图像上进行实验, 结果表明, 本文优化的Inception V3模型能够以71.5%的准确度对这些图像的状态进行分类, 与对比算法相比, 在分类准确度、训练损失上都有明显提升, 可以满足图像分类的可靠性、稳定性等要求。
Abstract
In order to realize the classification and recognition of object state, the Inception V3 module on the basis of GoogLeNet is optimized Using Tanh as the activation function and RMSprop, the SGD optimizer improves the accuracy of the model. Firstly, using cubic convolution interpolation, GAN preprocesses the image set. Then, the Inception is used to train the image. Finally, the SGD optimizer is combined with RMSprop to optimize the model. Experiments on the images of 20 cooking objects using the model proposed in this paper show that the optimized Inception V3 model can classify the state of these images with 71.5% accuracy. Compared with the comparison algorithm, the classification accuracy and training loss are obviously improved, which can meet the requirements of reliability and stability of image classification.
王旖旎. 基于Inception V3的图像状态分类技术[J]. 液晶与显示, 2020, 35(4): 389. WANG-Yi ni. Image classification technology based on inception V3[J]. Chinese Journal of Liquid Crystals and Displays, 2020, 35(4): 389.