基于改进DenseNet的刺绣图像分类识别的研究

Research on Embroidery Image Classification and Recognition based on improved DenseNet

摘要:

针对中华传统刺绣工艺传承保护问题中的分类任务，传统的刺绣分类方法存在耗时长、精度低以及需要大量掌握专业知识的人力资源等问题；设计了一种基于改进DenseNet的刺绣图像分类识别方法；构建刺绣图像分类识别数据集；采用局部二值模式LBP、Canny算子边缘提取以及Gabor滤波等方式提取纹理特征，将不同特征图与原图合并为四至六通道图像数据集送入网络进行消融试验，扩充了数据集宽度；为稳定训练过程，加速损失收敛速度，提出引入SPP (Spatial Pyramid Pooling)结构优化模型；为提高分类识别精度使用Leaky ReLU激活函数优化ReLU函数；实验结果表明基于改进DenseNet的刺绣图像分类识别方法可解决传统刺绣图像分类方法中存在的问题，改进后的刺绣图像分类模型与基准模型相比准确率提高了8.1%，高达97.39%。

Abstract:

Aiming at the problems of time-consuming and low accuracy of traditional embroidery classification methods in the inheritance and protection of traditional Chinese embroidery technology, an embroidery image classification method based on improved DenseNet is proposed. The local binary pattern, Canny operator edge extraction and Gabor filtering are used to extract the texture feature and the original image, which are merged into a four to six-channel image data set and sent to the network to expand the data set width. The Spatial Pyramid Pooling (SPP) structural optimization model is proposed to accelerate the convergence rate of loss. The Leaky ReLU activation function is used to optimize the ReLU function to improve the classification and recognition accuracy. The simulation results show that the embroidery image classification and recognition method based on the improved DenseNet can solve the problems existing in the traditional embroidery image classification method. The accuracy of the improved model is 8.1 % higher than that of the benchmark model, as high as 97.39 %.