A Segmentation Weighted Shoe Print Image Retrieval Method

Abstract

The present invention provides a kind of shoe print image retrieval method based on subsection weighting, comprising: step 1, image preprocessing step: based on the trained U-Net convolutional neural network model, the obtained shoe print scene photos are converted into reserved shoes On-site binary image of printed information; step 2, feature extraction step: split the on-site binary image and the image in the shoe sample library into upper and lower sub-images, input them into the twin network for feature extraction, and obtain two mutually independent sub-images Features; step 3, feature weight matrix calculation step: calculate the proportion of pixels containing shoe print information in two mutually independent sub-features, and thus obtain the weight matrix of the scene binary image; step 4, feature fusion and similarity degree measurement steps. The invention solves the technical problem of how to quickly and accurately retrieve the styles corresponding to the on-site photos of shoe prints in the shoe sample database, and comprehensively considers the characteristics of the large noise of the on-site shoe prints and the lack of partial information, and improves the accuracy of shoe print retrieval.

Description

technical field [0001] The invention relates to the technical field of retrieval, in particular to a method for retrieving shoe print images based on segment weighting. Background technique [0002] Shoeprints are one of the most common clues in crime scenes, and corresponding shoeprint identification is an important issue in forensic identification. The current shoe print recognition mainly has the following three challenges: (1) The shoe prints left on the scene contain complex environmental factors: according to the medium containing the shoe prints, they can be divided into leather, wood, soil, ceramics, etc. The production mechanism can be divided into two categories: additive and subtractive, such as Figure 1a shown. (2) The shoe prints left on the scene are often incomplete, so part of the information will be lost, such as Figure 1b shown. (3) The on-site shoe prints and the shoe prints in the shoe sample library come from different modalities. The former is a gr...

AI Technical Summary

Problems solved by technology

The current shoe print recognition mainly has the following three challenges: (1) The shoe prints left on the scene contain complex environmental factors: according to the medium containing the shoe prints, they can be divided into leather, wood, soil, ceramics, etc. The production mechanism can be divided into two categories: additive and subtractive, such as Figure 1a shown

(2) The shoe prints left on the scene are often incomplete, so part of the information will be lost, such as Figure 1b shown

Therefore, when it is transferred to the shoe print retrieval task, its network depth is small, and it is not easy to fit the distribution of complex shoe print data

Images