Object Recognition And Detection Based On Small Sample Learnin

Along with the increased performance of computing devices and the availability of very large datasets,deep learning has achieved excellent performance on data-intensive tasks.However,deep learning requires a large number of labeled data to train the model,and the training effect is not good on smaller datasets.Directly applying traditional deep learning methods,which require large amounts of data for training,to tasks with few labeled data is prone to overfitting,resulting in its performance in testing lower than the accuracy of training.In contrast,humans can quickly learn new concepts with only few data.Bridging the existed gap between deep learning and human learning has led to a boom in few-shot learning.To further improve the effectiveness of few-shot learning on object recognition and detection tasks,this thesis mainly does the following research work in terms of optimizing the representation of features:(1)Propose a few-shot classification model based on feature disentanglementIntroducing feature de-entanglement techniques to improve the autoencoder and using the optimized features for few-shot learning object recognition task.The designing of encoding and decoding modules based on feature disentanglement can ensure that the extracted semantic information is sufficient to be recovered into the original image.In addition,constraints are designed on the extracted features to ensure that the semantics of the same category are similar.Finally,use the semantic information to classification,and the experimental results are analyzed.(2)Propose a few-shot classification model based on feature differentiationExisting few-shot learning methods only view samples from an isolated perspective,ignoring the difference information between samples in the current scene.Therefore,a few-shot classification model based on feature differentiation is proposed.In order to make full use of information,this model proposes a novel feature adaptive fusion module,which fuses sample features under different transformations and mines samples’ information at different scales.In addition,this model designs a feature weighting module to mark semantic features with high discriminability,narrowing down the semantics of the same class and expanding the semantic gap between different classes.The experimental results demonstrate that this model can further learn how to distinguish different category concepts through differentiated features,with higher accuracy and robustness.(3)Propose a few-shot object recognition model based on attention mechanismIntroducing an attention mechanism module in the object detection task can help the model focus on the key information in the image,thus improve the model’s performance on the detection task.However,due to the limited data in the few-shot object detection task,the number of model parameters cannot be too large,so attention mechanism modules with fewer parameters need to be selected.A few-shot object detection model based on Mobile Vi T V2 Attention is proposed after considering the number of parameters and the final results.Experimental results on common datasets for object detection show that the proposed model can improve the accuracy on few-shot object detection tasks.In summary,in terms of optimizing the representation of features,this paper proposes a few-shot classification method based on feature disentanglement,a few-shot classification method based on feature differentiation and a few-shot object recognition method based on attention mechanism.These models are tested on publicly available datasets,and the results show that they can improve the accuracy of few-shot learning in object detection or recognition tasks.