Fine-grained Image Classification Combining Attention And Deep Learning

Fine-grained image classification,also called sub-category image classification,is a very popular research topic in the fields of computer vision in recent years.Its purpose is to make more detailed sub-categories of coarse-grained large categories.The classification accuracy of fine-grained images is more meticulous,and the differences between classes are more subtle.Often,only small local differences can be used to distinguish different classes.This paper applies a fine-grained image classification model that focuses on deep learning.The main work is as follows:1)A fine-grained image classification algorithm based on channel attention is applied.For high-resolution regions with fine-grained image classification tasks,it is not easy to receive too much attention,and it is vulnerable to background and external factors.By analyzing the generation form of the attention mechanism,the top-down attention is used to make the convolutional neural network have a predetermined purpose and task and focus on a certain object according to the object-level label information of the image.This method first uses the attention network to extract the advanced features of the image;then uses the regional suggestion network to locate the image target;and finally uses the Fast R-CNN network to classify and regression the target.Experimental results on the CUB-200-2011 data set and Stanford Dogs data set obtained 77.58% and 83.26% classification accuracy respectively.2)A fine-grained image classification algorithm with weakly supervised two-level attention is applied.Aiming at the problem that fine-grained image classification tasks require a lot of manual labeling costs,a fine-grained image classification method based on weakly supervised two-level attention is applied.This method first uses the R-CNN method and the global average pooling method on the object-level attention model to filter and locate saliency regions of the image to learn object features;then uses the neural network-based spectrum on the partial-level attention model The clustering pattern alignment method learns the subtle and local characteristics of the target;finally,a classification network will be used to combine the overall and local information of the target,and the final classification result will be obtained by superposition.Compared with other methods on the CUB-200-2011 dataset,this method achieved a classification accuracy of 87%.