Research On Fine-Grained Image Recognition Algorithm Based On Weak Supervision

In general image recognition tasks,images are often divided into rough categories in the traditional sense such as cats,dogs,and people,and their features are mostly based on rough outlines rather than details.In most actual businesses,it is often necessary to identify more fine-grained sub-categories of objects in images.However,in finegrained images,the representations of individuals of different categories are very similar,while individuals of the same category are very different due to reasons such as posture,growth period and angle.Coupled with object occlusion,complex background and other interference,its recognition difficulty is far more difficult than ordinary image recognition.In response to these problems,this paper starts from locating the discriminative area of the image to reduce the influence of intra-class differences,and extracting effective distinguishing features to increase the degree of discrimination of similar categories.And weak-supervised fine-grained image recognition algorithms are carried out.This paper designs a weakly-supervised fine-grained image recognition network based on auxiliary task,which simultaneously handles classification task based on multiscale feature and the weakly supervised detection auxiliary task.By introducing the region recommendation module,in the case of only image classification tags,multiscale and discriminative local area image blocks are recommended,and the multi-scale features of the image are extracted for fine-grained image classification.At the same time,it is proposed to add a parallel multi-branch weakly supervised detection module,only under the supervision of the classification label information,using the high-level semantic features of the convolutional neural network to train and detect the saliency distribution of the object.The complementary erasure training strategy and the crossgrouping convolution between the multi-branch channel features train the network to focus on the richer features of the object to be identified and reduce background interference.The network has achieved Top-1 accuracy rates of 88.0% and 92.8% on the CUB-200-2011 data set and FGVC Aircraft data set,which are 3.3% and 2.7% higher than the baseline method.Fine-grained recognition is more concerned about the discriminative local areas and fine feature expression than general image recognition.The part-based feature representation can not only enhance the discrimination between similar categories,but also effectively cope with intra-class differences caused by camera shooting angles,target poses,and appearance changes.This paper combines the results of weakly supervised detection based on high-level semantics and low-level features to refine object positioning,and trains a semantic-based part segmentation module that is independent of category.Without using segmentation tags,by combining three component segmentation standards,the part segmentation module is trained to obtain parts with geometric concentration,semantic consistency and classification discrimination.According to the spatial distribution information of different parts,the low-level convolution features are screened and aggregated to obtain part feature expressions,which enhance the difference of similar categories.At the same time,a weighted cross-entropy loss function is proposed to increase the model’s attention to the categories that are not easy to classify during training.Finally,the Top-1accuracy rates of 88.4% and 93.2% were achieved on the CUB-200-2011 data set and FGVC Aircraft data set.