CN105069144A - Similar image search method - Google Patents

Abstract

The invention discloses a method for searching similar pictures, comprising the following steps: S1 constructs a RootSIFT model of an image, and extracts RootSIFT features from images in an image database, the RootSIFT features include key points and feature descriptors, and extract The feature descriptors in the feature database are stored in the feature database; S2 extracts the RootSIFT features of the target image, uses the Flann feature matching method to match with the feature descriptors in the feature database, and calculates the number of key points that are successfully matched, that is, the distance between the two images distance; S3 outputs similar images.

Description

A way to search for similar images

technical field

The invention relates to the field of image processing, in particular to a method for searching similar pictures.

Background technique

With the development of digital information technology, digital images have entered thousands of households. However, with the increase in the number of digital images, how to quickly and accurately find other pictures similar to a certain picture in a huge image database has become a problem. problem. For example, a person traveled to Egypt, where the scenery is beautiful and the culture is profound, so he took thousands of photos. One day he occasionally found a photo of the pyramids at dusk that he particularly liked, and then he wondered if there were any photos of similar scenes in his camera. It would be a huge workload to browse through his photo collection one by one, Can you build a personal image search engine? As long as a certain picture is input into the engine, the image search engine will extract the features of the image, then match it with the image features in the image database, and finally output the image with the highest similarity with the input image.

SIFT is a scale-invariant feature transform (Scale-invariant feature transform, SIFT), which is a descriptor used in the field of image processing. This description has scale invariance, can detect key points in the image, and is a local feature descriptor. SIFT was proposed by David Lowe in 1999 and improved in 2004. The SIFT feature is based on some local appearance interest points on the object independent of the size and rotation of the image. The tolerance to light, noise, and micro viewing angle changes is also quite high. RootSIFT is an improved algorithm of SIFT. It uses the Hellinger kernel function instead of the Euclidean distance measure in SIFT as the similarity measure between two vectors. Experimental results show that the effect of RootSIFT is better than SIFT.

Contents of the invention

In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a method for searching similar pictures.

The present invention adopts following technical scheme:

A method for searching for similar pictures, comprising the following steps

S1 builds the RootSIFT model of the image, and extracts the RootSIFT feature to the image in the image database, the RootSIFT feature includes key points and feature descriptors, and stores the extracted feature descriptors into the feature database;

S2 extracts the RootSIFT feature of the target image, uses the Flann feature matching method, matches the feature descriptor of each image in the feature database, and calculates the number of key points that match successfully, that is, the distance between the two images;

S3 outputs the top n images with the largest number of key points that successfully match the target image.

The S1 is specifically:

S1.1 On the basis of SIFT, establish the RootSIFT model of the image;

S1.2 traverse each picture in the image database, and extract the RootSIFT feature of each picture through the RootSIFT model;

S1.3 Store the feature descriptor of the extracted RootSIFT feature into a pkl file through the pickle module of python.

The Flann feature matching method is specifically: calculating the distance between the feature descriptor vectors of two images, specifically calculating the distance between each key point that needs to be matched and each key point of the corresponding picture and finding the distance from the target key point. Key points for distance.

When the value of the closest distance divided by the next closest distance is less than 0.75, the two key points are considered to be matched successfully.

Beneficial effects of the present invention

(1) The present invention extracts the feature of image according to the improved algorithm RootSIFT feature extraction method of SIFT, and this feature has the invariance to rotation, scale, scaling, brightness change, describes the feature of image from different angles, can search and compare more accurately images similar to the target image;

(2) The present invention uses the Flann feature matching method during feature matching, so that the system has a fast matching rate and better search accuracy.

Description of drawings

Fig. 1 is a work flowchart of the present invention.

Detailed ways

The present invention will be described in further detail below in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example

As shown in Figure 1, a method for searching similar pictures includes the following steps:

S1 builds the RootSIFT model of the image, extracts the RootSIFT feature from the image in the image database, and stores the extracted feature descriptor in the feature database. The RootSIFT feature includes key points and feature descriptors, each key point has three information: position, scale, direction, a feature descriptor is established for each key point, and this key point is described by a set of vectors, Make it not change with various changes, such as lighting changes, viewing angle changes, and so on. The general feature descriptor calculates gradient information in 8 directions in a 4*4 window in the key point scale space, and a total of 4*4*8=128-dimensional vectors.

The RootSIFT model is extended from the SIFT feature model. The performance of using Euclidean distance in texture classification and image classification is lower than that of using Hellinger kernel function. Therefore, considering that the similarity measure in the SIFT algorithm can also be measured by the Hellinger kernel function, it is found that the effect of the kernel function is better.

Step S1 specifically includes:

S1.1 On the basis of SIFT, establish the RootSIFT model of the image.

The specific method is to use the Hellinger kernel function instead of the Euclidean distance to measure the similarity between two feature vectors. The specific operation can be divided into two steps: ① normalize the feature vector by L1, and ② calculate the square root of each element. Get the eigenvector RootSIFT with L1 normalization.

S1.2 traverse each picture in the image database, and extract the RootSIFT feature of each picture.

S1.3 Store the feature descriptor of the extracted RootSIFT feature into a pkl file through the pickle module of python.

S2 extracts the RootSIFT feature of the target image, uses the Flann feature matching method to match the feature descriptor of each image in the feature database, and calculates the number of key points that each image in the image database successfully matches the target image, that is Image-to-image distance.

The specific way is to use the Flann feature matching method, where k is set to 2, that is, the number of key points with the closest distance is selected as 2. Take a key point in an image and find the first two key points closest to Euclidean in another image. The specific calculation method is to calculate the Euclidean of the feature descriptor vectors of the two key points The distance is the sum of the squares of the difference between each corresponding element of the two vectors. Among these two key points, if the closest distance divided by the next closest distance is less than a certain ratio threshold, the pair of matching points is accepted. The reason for this operation is to exclude no matching due to image occlusion and background confusion. key points of the relationship. The threshold set in this method is 0.75. In this method, the number of key points that are successfully matched between two images is considered to be the distance between the two images. The more the number of successful matches, the closer the distance between the two images and the greater the similarity between the two images. high.

S3 outputs the top n images with the largest number of key points that successfully match the target image.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the embodiment, and any other changes, modifications, substitutions and combinations made without departing from the spirit and principle of the present invention , simplification, all should be equivalent replacement methods, and are all included in the protection scope of the present invention.

Claims (4)

1. A method for searching for similar pictures, characterized in that, comprising the steps of: S1 builds the RootSIFT model of the image, and extracts the RootSIFT feature to the image in the image database, the RootSIFT feature includes key points and feature descriptors, and stores the extracted feature descriptors into the feature database; S2 extracts the RootSIFT feature of the target image, uses the Flann feature matching method, matches the feature descriptor of each image in the feature database, and calculates the number of key points that match successfully, that is, the distance between the two images; S3 outputs the top n images with the largest number of key points that successfully match the target image. 2. A method for searching similar pictures according to claim 1, wherein said S1 is specifically: S1.1 On the basis of SIFT, establish the RootSIFT model of the image; S1.2 traverse each picture in the image database, and extract the RootSIFT feature of each picture through the RootSIFT model; S1.3 Store the feature descriptor of the extracted RootSIFT feature into a pkl file through the pickle module of python. 3. The method according to claim 1, wherein the Flann feature matching method is specifically: calculating the distance between the feature descriptor vectors of two images, specifically the distance between each key point that needs to be matched and the corresponding picture Each key point calculates the distance and finds the key point with the shortest distance from the target key point. 4. The method according to claim 1, characterized in that, when the value of the shortest distance divided by the next shortest distance is less than 0.75, the two key points are considered to be successfully matched.