Weakly Supervised Fine-grained Image Classification Based On Salient Region Location

In the field of computer vision,the samples to be classified in image classification tasks usually come from different basic categories(such as cars,dogs,birds,trees,etc.),however,in many practical application scenarios,these basic categories need to be further classified.The granularity of this classification is more detailed than that of general classification tasks,so it is called fine-grained image classification.Because fine-grained image classification tasks are classified into sub classes under the same basic category,the differences between sub classes are much smaller than those between the basic categories,and often reflected in small parts.This is the difficulty of fine-grained classification task.At the same time,the interference of image noise,shooting angle and lighting further increases the difficulty of fine-grained classification.This paper discusses the weak supervision classification technology based on the significance region.Firstly,it proposes a feature mining module which can make the location of the significant area more accurate by mining the dependency between channel features.Secondly,it attempts to propose a feature detection module which replaces RPN to generate the region of interest in the non-target detection field.In order to verify the effectiveness of the two methods,an existing NTS net based on the weak monitoring finegrained classification model based on the significant region is selected.Firstly,the two methods proposed in this paper are combined with the model,and then combined with the model.The experiments are carried out on the breast ultrasound data set and bird data set cuba-200-2001,The experimental results show that the two methods can improve the performance of the original model and be more beneficial to the classification of breast ultrasound.Because of the advantages of measurement learning in image classification,this paper also proposes a weak supervision and fine-grained classification network based on the significance region positioning based on metric learning.The network includes two symmetrical branch networks and measurement networks with shared weights,each of which includes convolutional neural network,feature mining module and feature detection module to extract image features,The similarity between pairs of learning samples is measured.In order to reduce the cost of measurement learning and testing,a classifier is trained to predict the test samples while training the measurement network.The network model is used in breast ultrasound data set.The experimental results show the effectiveness of the fine-grained classification method proposed in this paper.