Fine-grained image recognition algorithm for distributed labels based on inter-class similarity

Abstract

The invention discloses a fine-grained image recognition algorithm of distributed labels based on inter-class similarity. The fine-grained image recognition algorithm comprises the following steps: (1) extracting feature representation of an input image by using a backbone network; (2) calculating center loss and updating a category center by utilizing a center loss module through feature representation; (3) calculating, by a classification loss module, classification loss (such as cross entropy loss) by using the feature representation and final label distribution, wherein the final label distribution is obtained by calculating the weighted sum of one-hot label distribution and distributed label distribution generated by a category center; and (4) performing weighted summation on the center loss and the classification loss to obtain a final target loss function so as to optimize the whole model. The problem of overfitting is effectively relieved by reducing the certainty degree of model prediction, discriminative features of fine-grained data can be accurately learned, data of different fine-grained categories can be accurately and efficiently distinguished, and the invention can be widely applied to the fields of visual classification and multimedia.

Description

technical field [0001] The invention relates to a fine-grained image recognition method, in particular to a fine-grained image recognition algorithm based on distributed labels between classes. Background technique [0002] Image recognition, which aims to classify a given image, is a core research topic in the field of computer vision. General image recognition tasks (such as distinguishing airplanes, ships, dogs, cats, animals, etc., distinguishing different handwritten digits, distinguishing various objects of different classes, etc.) aim to classify and recognize images of different categories. Fine-grained image recognition, as an important and challenging subclass of image recognition, has been an active research field in computer vision. The goal of fine-grained recognition tasks is to identify different subcategories under the same category (such as distinguishing different categories of birds, airplanes, cars, etc.). In real life, fine-grained image recognition ha...

AI Technical Summary

Problems solved by technology

However, network models suitable for general image recognition (coarse-grained image recognition) tasks cannot achieve ideal results in fine-grained recognition tasks, mainly because of the following two factors: (1) Small differences between image classes: due to the fine-grained Categories belong to the same large category (for example, Acura RL Sedan 2012 and Buick Verano Sedan 2012 both belong to the parent category of car), so images of different categories tend to have similar features, which leads to high similarity between fine-grained categories (2) Large differences within image classes: Due to the differences in illumination, angle, occlusion, and even the parameter performance of the acquisition equipment during the image acquisition process, images of the same class often have large visual differences, especially the similarity between fine-grained classes The existence of high degrees makes the large difference between images in the same category become particularly obvious in fine-grained recognition tasks

Images