| Fine grained image recognition aims to identify different subcategory images.Compared to traditional image recognition tasks,the feature differences between different categories of targets to be recognized in this task are usually smaller,while the feature differences between different samples in the same category are significant.Considering the particularity of fine-grained recognition tasks,it is necessary to simultaneously locate and characterize the saliency local regions of the target in deep network design.At present,fine-grained image recognition has been widely applied in many fields such as intelligent driving,retail product production and processing,and ecological protection.When collecting real-world fine-grained images,it is unavoidable to face incomplete information samples.The main causes of incomplete information of the fine-grained targets to be recognized are obscuration between targets and non-targets or occlusion of key information due to the shooting angle.In this case,the occluded image blocks can cause interference to the attention of the shallow network of the fine-grained image recognition model,leading to the accumulation of saliency region localization errors layer by layer and eventually seriously affecting the recognition accuracy.Therefore,effectively locating and characterizing the unobscured information of the target becomes the key to summarize the fine-grained recognition accuracy of incomplete information samples.In addition,considering many practical application scenarios,there are harsh limitations on the realtime and storage space of models.Therefore,the study of lightweight models is crucial.To address the above application challenges and needs,the main innovations of this thesis are as follows:(1)Inspired by self-regulated learning in cognitive science,a self-regulated optimization model based on cyclic feedback multi-instance learning is proposed at the theoretical level.The innovation of the proposed optimization model lies in defining a single image as an example package composed of image blocks.By establishing a self feedback optimization objective function and simulating human progressive self-regulation behavior,the network can use the information in the deep network to correct shallow attention errors,thereby guiding the network to gradually focus on areas that are not occluded and contribute greatly to classification accuracy,significantly improving the recognition accuracy of images.(2)At the practical application level,we propose the SRLformer(Self-Regulated Learning Transformer)model based on single-scale global association by applying the tuning optimization model to the Transformer model.The innovation of the proposed model lies in the establishment of a closed-loop self attention mechanism for the first time.The SRLformer model consists of a forward self-attention branch and a backward self-attention branch,forming a mutually reinforcing closedloop feedback as a whole.The backward self-attention branch can periodically sparse the forward branch self-attention,which facilitates the filtering of image blocks that contain important semantic information.In addition,in order to ensure that the reverse branch generates positive gain,multiinstance aggregation feedback and sparse attention mask generation models are designed,both of which can effectively utilize the complementarity of forward and backward attention to ensure the accuracy of important image blocks localization.The experimental results show that the accuracy of the SRLformer on occluded datasets exceeds that of models such as Swin Transformer by up to 3.51%.(3)A lightweight self-regulating Transformer is proposed to effectively reduce the computational complexity of the SRLformer.The innovation of the proposed lightweight model lies in solving the dynamic one-to-many decision-making problem faced by feedback and block attention during the lightweight process.By simplifying the block attention mechanism and establishing a hierarchical structure,the complexity of the model is reduced.The experimental results show that the lightweight model can reduce the number of parameters by about 23% while only reducing the accuracy by about15%. |
