基于轻量化区域置信网络的细粒度图像分类

Lightweight regional confidence network for fine grained image classification

为提高细粒度图像分类的准确率和速度,提出区域投票分类模型和区域置信度机制以及基于轻量化区域置信网络的细粒度图像分类方法。将轻量化卷积神经网络分类器替换为区域投票分类器,加入区域置信机制,增加分类网络对于关键特征分类的权重,提升轻量化模型的准确率。在Cub200-2011数据集上的实验结果验证了区域投票模型和区域置信机制的有效性。相较于其它主流细粒度图像分类算法,改进后的模型仅损失了少量精度,却大幅减少了参数量和所需运算资源。

To improve the accuracy and speed of fine-grained image classification,the regional voting classification model,regional confidence mechanism and fine-grained image classification method based on lightweight regional confidence network were proposed.The lightweight convolutional neural network classifier was replaced by the region voting classifier.For improving the accuracy of the lightweight model,the region confidence mechanism was added to increase the weight of the classification network for the key feature classification.Experimental results on the Cub200-2011 dataset verify the effectiveness of the regional voting model and the regional confidence mechanism.Compared with other mainstream fine-grained image classification algorithms,the improved model only loses a small amount of accuracy,but greatly reduces the amount of parameters and computing resources.