Research On Fine-Grained Recognition Based On Distinctive Part

With the development of science and technology,people are increasingly demanding the classification of images.The traditional coarse-grained recognition become increasingly unable to meet the needs of classification tasks.In recent years,Fine-Grained recognition has attracted more and more attention in the fields of computer vision and pattern recognition.We studied the fine-grained recognition problem of weak supervision from three aspects: the priority selection of distinctive part,the expression of distinctive part,and the similarities between subcategories.The main research work is as follows:(1)Aiming at the problem that existing algorithms do not make good use of the similarity between subcategories,we propose a Multiple classifier based on similarity and difference.This method consists of two parts: S-SVM and D-SVM.We use S-SVM to learn the similarities between subcategories,and then use the similarities obtained to modify the SVM's multiclass function.Since the similarity may be formed by different factors(color of wings,shape of mouths),we use D-SVM to learn the differences between similarities of subcategories and use the obtained distinctiveness to modify the multi-classification function of SVM.Finally,we combine two modifications to get the SD-SVM multiple classifier.The experimental results on the CUB-Birds bird dataset and the FGVC-Aircraft aircraft dataset show that the proposed SD-SVM multiple classifier can effectively improve the accuracy of fine-grained recognition tasks.(2)Aiming at the lack of relationship description of distinctive part in existing fine-grained recognition.this paper proposes a multi-scale feature representation method based on distinctive part.First of all,through theoretical analysis and simulation experiments of multi-scale distinctive part of variance and variance,we demonstrate that there is a progressive relationship between the inter-regional classification capabilities as the scale changes.Based on this,the method uses RA-CNN to obtain distinctive part at three scales.According to the order of the scale from the largest to the smallest,the distinctive part convolution features are input to the recursive neural network to obtain the final features.Finally,Softmax is used to classify.The experiments on the CUB-Birds bird dataset show that the multi-scale distinctive part feature expression model is superior to the simple splicing method for distinctive part features.(3)Aiming at the problem that most existing methods cannot obtain reliable distinctive part,we propose a priority selection algorithm of distinctive part.Our method first uses the attention model to obtain candidate distinctive part,then calculate the geometric relationship between the candidate areas,remove the area where the position is abnormal,according to the calculated differences in the classification ability of different regions for different subcategories,unnecessary regions are removed,finally,the preferred regions are classified using a convolutional neural network.Our proposed algorithm achieves a correct rate of 87.0% in the CUB-Birds bird dataset and 90.2% in the FGVC-Aircraft aircraft dataset,reaching the current state-of-the-art level.