Method for retrieving footprint images

A method for retrieving footprint images, comprising:

• firstly, pre-training models through ImageNet data;
• cleaning footprint data and conducting expansion pre-processing for the footprint data by using the models which are pre-trained, and dividing the data sets of footprint images into five parts of a gallery data set, a query data set, a train data set, a train all data set and a val data set, wherein the query data set is formed by randomly selecting the images from each category in the gallery data set, each category in the gallery data set comprising more than 6 images similar to an image to be retrieved;
• the val data set is formed by selecting one footprint image from each category of the train all data set and normalizing the footprint data, and for the data sets, executing a normalization processing by a per-example mean subtraction method;
• adjusting full connection layers and classification layers of the models, comprising:
• modifying the full connection layers of three different models and subsequent parts of the full connection layers, and defining new full connection layers and new classification layers in an order of linear layers, batch standardization, linear rectification functions and linear layers, wherein the three different models are Resnet50 model, densenet121 model and VGG19 model;
• in the Resnet50 model and the densenet121 model, modifying original full connection layers to 512 dimensions and adding a new classification layer;
• in the VGG19 model, reserving a first full connection layer, removing a second full connection layer, adding a new classification layer, and determining the number of the new classification layers according to an image category in a training set initializing parameters of layers which are newly added by Kaiming initialization, wherein other parameters are from the models which are pre-trained on ImageNet;
• conducting training by the three adopted models on the data sets of the footprint images;
• training the models again by using the data sets through parameters of the models which are pre-trained;
• then, saving the models trained twice, removing the classification layer of the models trained twice, and executing a feature extraction for images in an image library and a retrieval library to form a feature index library;
• extracting features by three models, connecting the features extracted by the three models to form fused features, and establishing a fused feature vector index library;
• extracting the features of the images in the image library to be retrieved in advance, and establishing a feature vector library;
• and calculating distances in the retrieval library and the image library when a single footprint image is inputted, thereby outputting the image with the highest similarity.