Junk image fine-grained classification method based on incremental learning

Abstract

The invention relates to a junk image fine-grained classification method based on incremental learning. The junk image fine-grained classification method comprises the following steps: step 1, constructing a new and old category junk image database; step 2, respectively training a deep convolutional feature extraction network and an incremental classifier: firstly, training a resnet18-based deep convolutional neural network, called a resnet18 network, by utilizing the selected old category garbage image data set; removing a full connection layer from the trained resnet18 network to serve as adeep convolution feature extraction network of incremental learning; and finally, using the deep convolution feature extraction network to extract the deep convolution features of the old type of junkimages; extracting a deep convolution feature of a newly-added class garbage image by using a deep convolution feature extraction network as a negative class sample data set of an incremental SVM classifier, extracting a deep convolution feature of the newly-added class garbage image by using the deep convolution feature extraction network as a positive class sample data set of the incremental SVM classifier, and training the incremental SVM classifier; and step 3, establishing a classification incremental learning model.

Description

technical field [0001] The invention belongs to the field of image classification, and relates to a method for fine-grained classification of garbage images that is continuously and rapidly increasing by using a deep convolutional neural network and an incremental learning strategy. Background technique [0002] Image classification is one of the most basic tasks in the field of computer vision, and related technologies are widely used in smart cities, medical diagnosis, meteorological analysis, financial services and other industries. In recent years, with the rapid development of deep learning technology, image classification based on deep learning has also made new breakthroughs. From VggNet in 2014, stacked small convolution kernels are used [1] , In 2015, GoogleNet proposed inception to broaden the network width [2] , ResNet proposes cross-connection [3] , by 2017 SeNet won the last ImageNet recognition competition champion [4] , the final image classification TOP-5...

AI Technical Summary

Problems solved by technology

However, methods based on model parameter retention and other methods have poor classification accuracy on new categories. Although methods based on example sets have high accuracy, they need to reserve a large amount of storage space in advance. At the same time, it can realize the method of quickly identifying a small number of new categories, and solve the fine-grained classification of garbage images

Images