Research on Error Label Screening Method Based on Convolutional Neural Network
Zhengwen Li, Wenju Du, Nini Rao · 2021 IEEE 6th International Conference on Signal and Image Processing (ICSIP) · 2021
Supervised learning methods require a large number of labeled image data sets, but a large number of labeled image data sets are difficult to obtain in many practical applications, so it is necessary to develop weakly supervised learning methods. There is no good method for cleaning existing data sets containing false images that can simultaneously filter images with label noise and background noise. This paper presents a new method to solve the problem of image with label noise and background noise in weakly supervised learning, which can clean the wrong image with label noise and background noise. We first used the existing neural network VGG16_BN to extract the features of the image and delete the full connection layer. Then we added an error image automatic screening module to filter out the poor quality images and possible wrong images in priority. The images screened in this part include images with 40%-60% label noise and background noise. It can effectively reduce the screening range of the wrong image. We verified the effectiveness of the proposed method in the dataset of cat and dog dichoromy in the public dataset. We screened and determined that the dataset contained about 0.8% error images, and designed 200 error images to be screened by the proposed method. Finally, 186 designed error labels were screened out, which verified the effectiveness of the proposed method.