基于3D注意力的MobileNet图像分类算法改进

针对MobileNetV2网络在图像分类任务中特征表达不足的问题，提出一种结合注意力机制对MobileNet网络的改进策略。利用一种新颖的高效且无参的注意力模块，同时结合I-block模块来替换MobileNet网络中的倒残差模块，采用RReLU激活函数替代原ReLU激活函数保留更多特征，结合inception结构进行多尺度特征提取与融合，使其可以提供更强的多尺度特征表达并服务于图像分类任务，使用数据扩增技术，生成更多样本。与6种方法进行对比，实验结果表明，采用3D注意力机制的网络在数据集CIFAR-10、CIFAR-100上以最少的网络参数分别取得94.09%和75.35%的最高精度，表明该改进方法可以有效地进行快速图像分类。

New MobileNet image classification algorithm based on 3D attention

To address the problem of insufficient feature representation in image classification tasks in MobileNetV2 network, an improvement strategy combined with attention mechanism is proposed. A novel, efficient and parameterless attention module is used to replace the reciprocal residual module in the MobileNet network with the I-block module. The RReLU activation function is used to replace the original ReLU activation function to retain more features. The multi-scale feature extraction and fusion are combined with the inception structure, so that it can provide stronger multi-scale feature expression and serve image classification tasks. Data amplification technology is used to generate more samples. Compared with six methods, experimental results show that the network with 3D attention mechanism achieves the highest accuracy of 94.09% and 75.35% on datasets CIFAR-10 and CIFAR-100 with the least number of network parameters respectively, indicating that our improved method can be effective for fast image classification.