Research on Small Sample Bloody Violence Image Recognition Algorithm Based on Transfer Learning
Currently,traditional neural network method cannot quickly and accurately recognize the bloody and violent images when the samples are insufficient or uneven distributed.To solve this problem,an ECA-FOCALLOSS-RESNET50 small sample bad pictures recognition algorithm based on transfer learning is proposed.Firstly,the lightweight and efficient channel attention mechanism ECA module is introduced,which avoids the feature loss caused by the dimensionality reduction of the traditional atten-tion mechanism under the premise of involving a small number of parameters.This method has obvious effect gain with insufficient training samples.Secondly,by improving the loss function,the problem of uneven samples of bad images with small samples is effectively solved.Finally,pre-training is introduced by transfer learning,and parameters fine-tune on the self-built data set to alleviate the overfitting phe-nomenon caused by insufficient samples.The results show that the transfer learn-based ECA-FOCAL-LOSS-RESNET50 small sample bad pictures recognition technology can improve the recognition accuracy by 2.92％and the AUC value by 0.023 1 on the self-built data set,which provides a technical method for maintaining the clean Internet.
NETL
NSTL
万方数据
