CN102855473B - A kind of image multi-target detection method based on similarity measurement - Google Patents

Abstract

The invention relates to an image multi-target detection method based on similarity measurement. The specific steps of the method are as follows: step 1: extract the feature points of each image in the target sample library, and calculate the BOF word bag vector of each image, thereby forming the The BOF vector set of the target sample library; Step 2: Utilize the BOF vector set of the target sample library to construct the M-Tree index of the target sample library; Step 3: Carry out multi-target detection on the image to be detected in the M-tree index, And return the detection results to the user; this method uses BOF vectors for target detection, which has good robustness and is not sensitive to target deformation, occlusion, and illumination changes; multiple targets can be detected at one time, without repeated detection, and can be realized Efficient online multi-target detection; this method can be widely used in many fields such as network information security, image multi-target detection, and video multi-target tracking.

Description

A method for image multi-target detection based on similarity measure

technical field

The invention relates to an image detection method, in particular to an image multi-target detection method based on similarity measure.

Background technique

Image object matching and detection has always been a very important field in computer vision. In the detection of image objects, the detection methods mainly include detection methods based on local features of images and detection methods based on sliding windows. Wherein, the detection method based on local image features includes the following steps: feature detection, feature recognition, and model matching. However, the detection method based on the sliding window scans the input image through windows of different sizes, and judges whether the scanned image block contains the target object.

These algorithms are relatively mature and can achieve real-time detection, but these methods have scalability problems when applied to multi-target detection. When the detection method based on image local features detects multiple targets, it needs to match the feature points of each target image with the feature points of the image to be detected, and find the matching position of each target respectively, which will cause the matching time to vary with the target The increase of the number shows a linear increase trend; when the detection method based on the sliding window detects multiple targets, it needs to match the sub-blocks of the image to be detected with each target image to judge whether it is the target position, so the matching time will also decrease increases linearly with the number of targets.

From the above, the detection time of existing detection methods based on local image features and sliding window detection methods will increase linearly with the increase in the number of sample targets to be detected. For a large number of image objects, the detection efficiency is very low, and it is not suitable to apply the above methods. During the detection process, the user hopes that the system can have a good detection speed and can quickly detect the target in large-scale target image data.

Contents of the invention

The technical problem to be solved by the present invention is to provide an image multi-target detection method based on similarity measurement, which can detect multiple targets at one time, and is efficient, fast and accurate.

The technical solution of the present invention to solve the above-mentioned technical problems is as follows: an image multi-target detection method based on similarity measurement, characterized in that the specific steps of the method are as follows:

Step 1: Extract the feature points of each image in the target sample library, and calculate the BOF bag-of-words vector of each image, thereby forming the BOF vector set of the target sample library;

Step 2: Utilize the BOF vector set of the target sample library to construct the M-Tree index of the target sample library;

Step 3: Perform multi-target detection on the image to be detected in the M-Tree index, and return the detection result to the user.

The beneficial effect of the present invention is: a kind of image multi-target detection method based on similarity measure described in the present invention, this method utilizes BOF vector to construct M-Tree index to carry out target detection, has very good robustness, and the target deformation , occlusion, and illumination changes are not sensitive; multiple targets can be detected at one time without repeated detection, and efficient online multi-target detection can be realized; this method can be widely used in network information security, image multi-target detection, video multi-target tracking, etc. fields.

On the basis of the above technical solutions, the present invention can also be improved as follows.

Further, the extraction of the feature points of the image is carried out by using the SIFT scale-invariant feature transformation algorithm.

Further, the steps of calculating the BOF vector of each image are as follows:

1) Select a representative image library, and calculate the cluster centers of all feature points in the image library, that is, the feature dictionary;

2) Calculate the Euclidean distance between the feature points of the image of the BOF vector to be calculated and the cluster centers described in step 1);

3) Find the nearest class center from the feature point of the image to be calculated BOF vector, calculate the frequency of the class center, and add 1 to the frequency;

4) Repeat steps 2) and 3) until the frequency calculation of all the feature points of the image to be generated BOF vector is completed, and the frequency histogram of the image is obtained, and the frequency histogram is vectorized to obtain the BOF vector of the image .

Further, the described image to be detected is carried out multi-target detection in the M-tree index, and the steps are as follows:

1) Define a fixed-size sliding window as needed;

2) The sliding window slides on the image to be detected according to certain rules to obtain image blocks to be detected;

3) calculating the BOF vector of the image block to be detected;

4) matching the BOF vector of the image block to be detected in the M-Tree index;

5) Using a similarity measurement method to detect whether there is an object similar to the image block to be detected in the M-Tree index, if there is, perform step 6); otherwise, return to step 2);

6) Return the detected similar target images and their positions in the large image to be detected to the user.

Further, the certain rule is to slide from left to right and from top to bottom, and the distance of each slide to the right or down is a constant.

Description of drawings

Fig. 1 is the overall flow chart of a kind of image multi-target detection method based on similarity measure that the present invention relates to;

Fig. 2 is the flow chart of calculating the BOF vector step of each image involved in the present invention;

Fig. 3 is the flow chart of the multi-target detection steps of the image to be detected in the M-tree index involved in the present invention;

Fig. 4 is a schematic diagram of the sliding window sliding on the image to be detected involved in the present invention.

In the accompanying drawings, the list of parts represented by each label is as follows:

1. The image to be detected, 2. The sliding window.

detailed description

The principles and features of the present invention are described below in conjunction with the accompanying drawings, and the examples given are only used to explain the present invention, and are not intended to limit the scope of the present invention.

As shown in Figure 1, it is a general flow chart of a kind of image multi-target detection method based on similarity measure that the present invention relates to; Figure 2 is a flow chart of calculating each image BOF vector step involved in the present invention; Figure 3 is the present invention The flow chart of the multi-target detection steps in the M-tree index of the image to be detected involved; FIG. 4 is a schematic diagram of the sliding window sliding on the image to be detected involved in the present invention. As shown in Figures 1, 2, 3, and 4, the specific steps of an image multi-target detection method based on similarity measurement are as follows:

The user provides the target sample library and the image to be detected.

Step 1: extract the feature points of each image in the target sample library, and calculate the BOF word bag vector of each image, thereby forming the BOF vector set of the target sample library; the steps of calculating the BOF vector of each image are as follows:

1) Select a representative image library, and calculate the cluster centers of all feature points in the image library, that is, the feature dictionary;

2) Calculate the Euclidean distance between the feature points of the image to be calculated BOF vector and the cluster center described in step 1);

3) Find the nearest class center from the feature point of the image to be calculated BOF vector, calculate the frequency of the class center, and add 1 to the frequency;

4) Repeat steps 2) and 3) until the frequency calculation of all the feature points of the image to be generated BOF vector is completed, and the frequency histogram of the image is obtained, and the frequency histogram is vectorized to obtain the BOF vector of the image .

Step 2: Utilize the BOF vector set of the target sample library to construct the M-Tree index of the target sample library;

Step 3: Perform multi-target detection on the image to be detected in the M-tree index, and return the detection result to the user; the specific steps of detection are as follows:

1) Define a fixed-size sliding window 2 as required;

2) The sliding window 2 slides on the image to be detected 1 according to certain rules to obtain the image block to be detected;

3) calculating the BOF vector of the image block to be detected;

4) matching the BOF vector of the image block to be detected in the M-Tree index;

5) Using a similarity measurement method to detect whether there is an object similar to the image block to be detected in the M-Tree index, if there is, perform step 6); otherwise, return to step 2);

6) Return the detected similar target images and their positions in the large image to be detected to the user.

Wherein, the certain rule is to slide from left to right and from top to bottom, and the distance of each slide to the right or down is a constant.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within range.

Claims (4)

1. based on an image multi-target detection method for similarity measurement, it is characterized in that, the method concrete steps are as follows:

Step one: the unique point extracting every width image in target sample storehouse, calculates the BOF word bag vector of every width image, thus forms the BOF vector set in described target sample storehouse;

Step 2: utilize the set of the BOF in described target sample storehouse vector to build M-tree index in described target sample storehouse;

Step 3: image to be detected is carried out multi-target detection in M-tree index, and testing result is returned to user;

Described image to be detected is carried out multi-target detection in M-tree index, and testing result is returned to user, step is as follows:

1) moving window of a fixed measure is defined as required;

2) described moving window slides according to certain rule on image to be detected, obtains image block to be detected;

3) the BOF vector of described image block to be detected is calculated;

4) BOF of described image block to be detected vector is mated in described M-tree index;

5) utilize method for measuring similarity to detect in described M-tree index whether there is the target similar to described image block to be detected, if existed, perform step 6), otherwise, return step 2);

6) the similar purpose image detected and its position in image to be detected are returned to user.

2. a kind of image multi-target detection method based on similarity measurement according to claim 1, it is characterized in that, the extraction of described image characteristic point utilizes sift Scale invariant features transform algorithm to carry out.

3. a kind of image multi-target detection method based on similarity measurement according to claim 1, is characterized in that, the BOF vector of the every width image of described calculating, and step is as follows:

1) choose a representative image library, and calculate the cluster class center of all unique points of this image library, i.e. feature lexicon;

2) calculate unique point and the step 1 of the image of BOF vector to be calculated) described in cluster class center between Euclidean distance;

3) find from the nearest class center of the unique point of the image of BOF vector to be calculated, calculate the frequency at such center, and frequency is added 1;

4) repeat step 2), 3), until complete all unique points of the image of BOF vector to be generated frequency calculate, obtain the frequency histogram of this image, namely described frequency histogram vectorization is obtained this image BOF vector.

4. a kind of image multi-target detection method based on similarity measurement according to claim 1, is characterized in that, described certain rule is according to from left to right, slides in top-down direction, at every turn to the right or the distance of slide downward be a constant.