Research On Cross-Domain Few-Shot Image Recognition Based On Pseudo-Siamese Neural Network Structure

Cross-domain few-shot learning is regarded as one of the research highlights in machine learning recently.The difficulty of the research lies in the accuracy drop of crossdomain network learning on a single domain due to the differences between the domains.To alleviate the problem,according to the idea of contour cognition and the process of human recognition,a few-shot learning method based on pseudo-Siamese convolution neural network is proposed.The original image and the sketched map are respectively sent to the branch network in the pre-training and meta-learning process.While maintaining the original image features,the contour features are separately extracted as branch for training at the same time to improve the accuracy and generalization of learning.The main work of this paper is as follows:(1)A sample preprocessing method based on image edge detection is proposed,which extracts the contour features of the sample,and sends the original sample and the extracted features to the two branches of the neural network respectively.And the inference that the contour feature is one of the key points of object classification is verified experimentally.The results show that the neural network can classify through the contour feature,which is less affected by the difference between the domains than the original image.(2)The existing cross-domain few-shot learning methods is analyzed,and a crossdomain few-shot learning method based on the pseudo-Siamese neural network structure is proposed.In order to preserve the recognition accuracy of the source domain while enhancing the generalization ability,the pseudo-Siamese convolutional neural network structure is introduced to deal with the cross-domain few-shot problem while large differences exist between the source domain and the target domain.While keeping the original feature extraction process unchanged,a branch network that takes the extracted contour feature as input is added to improve the generalization ability.At the same time,the similarity measurement module of the pseudo-Siamese neural network is modified,so that the network can adapt to the classification task.(3)Interpretability has always been a hot and difficult spot in the research process of deep learning.Due to the large amount of deep learning parameters and uncertainties,the order of input samples also has a certain impact on the final result of training.This paper also explores the interpretability.For the network proposed in this paper,a heat map is generated for the network after pre-training and meta-learning process,and the heat map is analyzed to intuitively illustrate the feasibility of the network.Cross-domain few-shot learning experiments are conducted and good results have been achieved by using mini-Image Net as source domain,Euro SAT and Chest X as the target domains.Also,the results are qualitatively analyzed using a heatmap to verify the feasibility of our method.