基于改进ShuffleNetV2的图像分类算法

Image Classification Algorithm Based on Improved ShuffleNetV2

摘要: 对ShuffleNetV2的轻量级分类网络进行了研究，提出了一种改进图像分类方法，来提高图像分类的准确性。相比于传统的卷积神经网络，可以提高模型的分类准确度和速度。首先，改进的模型引入6 × 6的深度分离卷积来替代原来3 × 3的深度分离卷积，来增加分类的准确度，其次，改进的模型应用了扁平层对数据进行降尺度和增加全连接层来提升学习能力。此外，改进的模型还对比了其他传统CNN模型的图像分类表现，选取cifar10/100数据集进行实验，实验结果表明改进的模型在准确度上提高了2%以上。

Abstract: This study investigates the lightweight classification network of ShuffleNetV2 and proposes an improved method for image classification to enhance accuracy. In comparison to traditional con-volutional neural networks, the improved model achieves higher classification accuracy and faster processing speed. Firstly, the enhanced model introduces 6 × 6 depthwise separable convolution to replace the original 3 × 3 depthwise separable convolution, thereby increasing classification accuracy. Secondly, the improved model incorporates a flatten layer for data dimensionality re-duction and adds a fully connected layer to enhance learning capabilities. Additionally, the im-proved model compares the image classification performance with other traditional CNN models, conducting experiments on the cifar10/100 datasets. The experimental results demonstrate an increase in accuracy of over 2% for the improved model.