A method for footprint image retrieval

Abstract

The invention relates to a method for footprint image retrieval. The method comprises the following steps: pre-training a model through ImageNet data; adopting the pre-trained model to perform cleaning and expansion preprocessing operation on footprint data, dividing the footprint data into a plurality of data sets with different purposes, adjusting a full connection layer and a classification layer of the model, adopting parameters of the pre-trained model, and using the footprint data sets to perform model retraining; storing the secondarily trained model, removing a classification layer ofthe model, and performing feature extraction on the images in the image library and the retrieval library to form a feature index library; connecting the features extracted by the three models to forma fusion feature, and establishing a fusion feature vector index library; extracting image features in a to-be-retrieved image library in advance, establishing a feature vector library, and when a single footprint image is input, carrying out distance calculation in the retrieval library and the image library, and outputting an image with the highest similarity. The footprint image recognition and retrieval functions are realized.

Description

technical field [0001] The invention relates to a method for footprint image retrieval. Background technique [0002] At present, the research on image retrieval began in the 1970s, and it was all based on text image retrieval at first, and described the features of images by using words. In the early 20th century, content-based image retrieval began to appear, that is, the color and texture of the image were analyzed, and shallow classifier techniques such as support vector machines were used to improve the accuracy of the search. But these methods still cannot solve the problem of semantic gap. With the research and development of deep learning, convolutional neural network [CNN] has performed well in image retrieval and recognition in recent years. With the holding of various types of image recognition competitions (such as ImageNet and Kaggle), various variant models of convolutional neural networks such as AlexNet[12], VGG, GoogLeNet, ResNet and DenseNet rely on their...

AI Technical Summary

Problems solved by technology

Traditional footprint retrieval and identification methods often consume a lot of human resources and time, and manual retrieval under large amounts of data is prone to errors

Therefore, the automatic footprint retrieval system has a strong application demand; however, in the research of the automatic footprint matching system, the traditional shallow machine learning method is often used to process the footprint image, which will take a lot of time, and the accuracy rate needs to be improved

Images