Research On Deep Multiple-instance Algorithms Based On Channel And Spatial Attention

With the rapid development of Internet technology,a large amount of data has been generated in related fields,and the computing power of hardware has also produced earthshaking changes,gradually making deep learning more developed in the field of artificial intelligence.However,many deep learning methods rely on a large amount of labeled data.The huge amount of data makes labeling time-consuming and laborious.To cope with this actual situation,weakly supervised learning has aroused strong interest among researchers.As an imprecise supervised algorithm in weakly supervised learning,multi-instance learning has been widely used in related fields such as computer vision and audio detection.This article will explore the multi-instance learning,and integrate it with deep learning,combining channel attention and spatial attention mechanisms,and propose two deep multi-instance learning algorithms based on channel attention and spatial attention.(1)A deep multi-instance network LBSN based on learnable similarity between bags is proposed.First,the model aggregates the new feature expression of reference bags by the weighted Att Net basic network that containing the attention layer,and aggregates the new feature expression of the target bag by initial Att Net that containing the attention layer;secondly,it calculates the similarity between the target bag and reference bags based on new feature expressions to obtain the similarity matrix;then,in order to solve the problem that the bags' similarity can be learned,the relationship between important bag and the relationship between important instances are learned through the channel attention and the spatial attention.The weighted similarity matrix will be formed a new inter-bag information vector and send the vector to the bag-level classifier to achieve classification,finally,a new LBSN trained by the classification loss.The design of this decoupling scheme reduces the time of training network.(2)A deep multi-instance network S-LBSN based on self-attention is proposed.First,the model uses the feature extraction module in the basic network Att Net to aggregate the new feature expression of the bag.The method adopted is to use the trained feature extraction module to obtain the new feature expressions of all reference bags,and then train a new network.This new network first contains the feature extraction module,but the parameters in the feature extraction module are not pretrained,and the target bag's new feature expression is obtained by the untrained feature extraction module,and secondly,the network calculates the similarity matrix between the new feature expression of the reference bags and the new feature expression of the target bag.Then,there will be two branches that use the self-attention mechanism to capture the internal relationship.The first branch is the channel self-attention to capture the internal relationship between bags,and the channel weight is assigned to the similarity matrix.The second branch is the spatial self-attention capture feature internal relations,and the spatial weight is assigned to the similarity matrix.Finally,the maximum response point strategy is adopted for the weighted similarity matrix,that is,a new bag-level feature vector is formed and sent it to the classifier to achieve classification.The above two algorithms are verified by many comparative experiments on a variety of multi-instance basic datasets.The experimental results show the effectiveness of the two algorithms and the modules added in this article.