Image Analysis Based On Distillation Learning

With the rapid development of deep learning and the emergence of more and more complex problems related to computer vision in the real world,the performance of image analysis models has been improved by leaps and bounds.At the same time,due to the rapid development of the neural network,the computational cost required by the model also shows an explosive growth.Therefore,it is of great benefit and value of research to reduce the parameters required by the model and improve the running speed by ensuring the performance of the model.Distillation learning is a model lightweighting method proposed in recent years,which usually extracts the knowledge of a large model and imparts it to a small model.However,due to the one-side traditional distillation learning knowledge transfer,the performance of small models depends entirely on the large model,which limits the performance of small models.And because the previous distillation models focus on the classification problems,it is difficult to solve the regression problems.Therefore,this paper improves the model and proposes an image analysis algorithm based on distillation learning.Firstly,for the analysis of multiple images in the dataset,in order to simplify the model and improve the performance of the teacher model and the student model,we propose a deep metric learning based on distillation learning.In this model,we use the advantages of the existing metric learning method and alleviate the over-fitting problem of the large model and improve the performance of the small model greatly by the guidance of the teacher model and the feedback of the student model.Secondly,for the image analysis of a single picture,we make improvements on the characteristics of traditional distillation learning which can only solve the classification problems and can not solve the regression problems.Combining the distillation model with the attention mechanism,we add supervised information to the small model in the high-level network of the large model.Meanwhile we add the feature adaptation layer to map features from the small network onto the same dimension of the large network.And by adding attention mechanism module we make the objects more concerned,reducing the interference of background noise information.We study the metric learning and object detection algorithms in image analysis,and reduce the complexity of the model through distillation learning.Experiments have been carried out on several public datasets.The ideal experimental results are obtained in both clustering algorithm and object detection problem,and the effectiveness of the proposed algorithm has been proved.