一直以来,由于大规模图像种类繁多且形态各异,导致大规模图像识别领域研究发展非常缓慢。在深度模型中,卷积神经网络(convolutional neural network, CNN)可以提取颜色、轮廓等浅层特征。随着层次的加深,其特征表述也由颜色、轮廓等特征逐渐抽象为整体特征。然而通过实验发现,网络的这种特征提取方式在提取整体特征时会出现一些不利于有效分类的“坏通道”。这种现象在大规模的图像分类任务中表现的更加明显。这些通道参与了网络的后续计算并且一定程度上降低了网络的性能。为了筛选出这些不利于分类的通道,提出了结合L1和L2范数进行特征选择的方法。通过对比多个网络模型的实验结果,该特征选择算法在大规模图像识别中具有更好的性能,并且可以提高网络的识别准确率。

For a long time, due to the wide variety and different forms of large-scale images, the research and development of large-scale image recognition has been very slow. In the deep model, convolutional neural network can extract shallow features such as color and contour. With the deepening of the level, the expression of its characteristics is gradually abstracted from the characteristics of color and outline to the overall characteristics. However, we found through experiments that this feature extraction method of the network will have some "bad channels" that are not conducive to effective classification when extracting overall features. These channels participate in the subsequent calculation of the network and reduce the performance of the network to a certain extent. In order to screen out these channels that are not conducive to classification, this paper proposes a method of combining L1 and L2 norms for feature selection. By comparing the experimental results of multiple network models, the feature selection algorithm has better performance in large-scale image recognition and can improve the recognition accuracy of the network.