摘要：Images play an important role in modern internet communications, but not all of the images shared by the users are appropriate, and it is necessary to check and reject the inappropriate ones. Deep neural networks do this task perfectly, but it may not be necessary to use maximum power for all images. Many easier-to-identify images may be classified at a lower cost than running the full model. Also, the pressure on the system varies from time to time, so an algorithm that can produce the best possible results for different budgets is very useful. For this purpose, a deep convolutional neural network with the ability to generate several outputs from its various layers has been designed. Each output can be considered as a classifier with its own cost and accuracy. A selector is then used to select and combine the results of these outputs to produce the best possible result in the specified time budget. The selector uses a reinforcement learning model, which, despite the time-consuming learning phase, is fast at execution time. Our experiments on challenging social media images dataset show that the proposed model can reduce the processing time by 32 % by sacrificing only 1.4 % of accuracy compared to the VGG-f network. Also, using different metrics such as F1-score and AUC (the Area Under the Curve in the accuracy vs. time budget chart), the superiority of the proposed model at different time budgets over the base model is shown. Download data is not yet available.
其他摘要：Images play an important role in modern internet communications, but not all of the images shared by the users are appropriate, and it is necessary to check and reject the inappropriate ones. Deep neural networks do this task perfectly, but it may not be necessary to use maximum power for all images. Many easier-to-identify images may be classified at a lower cost than running the full model. Also, the pressure on the system varies from time to time, so an algorithm that can produce the best possible results for different budgets is very useful. For this purpose, a deep convolutional neural network with the ability to generate several outputs from its various layers has been designed. Each output can be considered as a classifier with its own cost and accuracy. A selector is then used to select and combine the results of these outputs to produce the best possible result in the specified time budget. The selector uses a reinforcement learning model, which, despite the time-consuming learning phase, is fast at execution time. Our experiments on challenging social media images dataset show that the proposed model can reduce the processing time by 32 % by sacrificing only 1.4 % of accuracy compared to the VGG-f network. Also, using different metrics such as F1-score and AUC (the Area Under the Curve in the accuracy vs. time budget chart), the superiority of the proposed model at different time budgets over the base model is shown.
关键词：Adult content recognition; time-sensitive; cost-sensitive model; convolutional neural network; deep learning; image classification Abstract Images play an important role in modern internet communications; but not all of the images shared by the users are appropriate; and it is necessary to check and reject the inappropriate ones;Deep neural networks do this task perfectly; but it may not be necessary to use maximum power for all images;Many easier-to-identify images may be classified at a lower cost than running the full model;Also; the pressure on the system varies from time to time; so an algorithm that can produce the best possible results for different budgets is very useful;For this purpose; a deep convolutional neural network with the ability to generate several outputs from its various layers has been designed;Each output can be considered as a classifier with its own cost and accuracy;A selector is then used to select and combine the results of these outputs to produce the best possible result in the specified time budget;The selector uses a reinforcement learning model; which; despite the time-consuming learning phase; is fast at execution time;Our experiments on challenging social media images dataset show that the proposed model can reduce the processing time by 32 % by sacrificing only 1.4 % of accuracy compared to the VGG-f network;Also; using different metrics such as F1-score and AUC (the Area Under the Curve in the accuracy vs;time budget chart); the superiority of the proposed model at different time budgets over the base model is shown;Downloads Download data is not yet available.
