Sensitive Image Recognition based on Improved Channel Distillation Model

With the swift development of Deep Learning (DL) techniques, Deep Neural Networks (DNNs) have achieved remarkable performance in image recognition. However, they still encounter various challenges in sensitive image recognition. In the absence of extensive and well-annotated datasets, these methods may be prone to overfitting and could exhibit unstable output responses during the classification process. Moreover, the current mainstream deep learning models have a considerable number of parameters and are difficult to deploy on the embedded side. To solve this problem, this paper firstly conducts a comprehensive evaluation of three mainstream distillation model architectures, namely knowledge distillation, joint training, and channel distillation. Secondly, we further improve the channel distillation model, optimize the network structure, incorporate a random image generation approach to enhance the model's generalization capability, and employ a weighted cross-entropy loss function to boost the accuracy of complex image recognition. The experimental results indicate that channel distillation demonstrates the best recognition performance on sensitive image datasets. The proposed improvement strategy effectively reduces overfitting and improves the classification accuracy.