An anti-interference target acquisition method based on machine vision technology measurement

Abstract

The present invention discloses an anti-interference target acquisition method based on machine vision technology measurement, the steps of which include: S1, original image multi-stage filtering preprocessing; S2, target feature extraction; S3, random forest construction: calling a historical data sample library to obtain an anti-interference target acquisition library, randomly selecting M data points from it, creating a decision tree for the selected data points, each decision tree will produce a result, comprehensively analyze the results and output a majority vote or average; S4, through random forest, obtain a target feature reference set; S5, select the first N frames of the on-site detection picture as a training set, obtain appropriate counting weights through continuous training, and then start from the N+1th frame image, perform voting analysis by determining the weight and using the Adaboost algorithm to determine the final target. The present invention can effectively solve the interference problem in the target capture process under complex scenes, and can effectively improve the accuracy and robustness of target detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine vision, and in particular to an anti-interference target acquisition method based on machine vision technology measurement. BACKGROUND

[0002] Machine vision is an important branch of artificial intelligence; machine vision is to replace human eyes to do measurement and judgment, and the machine vision system is to convert the target to be taken into image signals through machine vision products (i.e. image taking device, such as industrial camera), and transmit the image signals to a special image processing system to obtain the shape information of the target to be taken, and convert the information such as pixel distribution and brightness, color into digital signals; the image system performs various operations on these signals to extract the features of the target, and then controls the action of the equipment on site according to the judgment result.

[0003] It should be pointed out that for the machine vision system, how to effectively process the image obtained by the machine vision instrument is the most important; among them, the existing technology mainly focuses on the recognition and detection of target image, such as Chinese invention patent with patent number ZL202310712312.4 and patent name Fish Swimming Speed Recognition Method Based on Machine Vision, Chinese invention patent with patent number ZL202311312489.1 and patent name Goose Shed Disinfection Liquid Spray Image Detection Method Based on Machine Vision, Chinese invention patent with patent number ZL202310124844.6 and patent name Seedling Fertilization Control System Based on Machine Vision, etc.

[0004] It should be emphasized that since the existing technology has not considered the influence of external interference factors on the detection and recognition of image target; in real use scenarios, the image acquisition process is often affected by external interference, if this problem is not solved, it may lead to target acquisition failure, making the subsequent image processing useless.

[0005] Therefore, in order to solve the problem that the target capture is easy to fail when facing external interference factors with complex gray scale, shape and dynamic characteristics in the scene captured by machine vision instrument, it is necessary to provide an anti-interference target acquisition method based on machine vision technology measurement. SUMMARY

[0006] The present application aims at the deficiencies of the prior art to provide an anti-interference target acquisition method based on machine vision technology measurement, which can effectively solve the interference problem in the target capture process in complex scenes, and can effectively improve the accuracy and robustness of target detection.

[0007] In order to achieve the above-mentioned purpose, the present application is realized by the following technical scheme.

[0008] An anti-interference target acquisition method based on machine vision technology measurement, comprising the following steps, in particular:

[0009] Step S1, image preprocessing: the captured original image is subjected to multi-stage filtering preprocessing;

[0010] Step S2, target feature extraction: the target in the image preprocessed in step S1 is subjected to feature extraction, and specifically includes shape feature extraction, radiation feature extraction and motion feature extraction;

[0011] Step S3, random forest construction: a historical data sample library is called to obtain an anti-interference target acquisition library, M data points are randomly selected therefrom, a decision tree is created for the selected data points, each decision tree produces a result, and a majority vote or average is output after comprehensive analysis of the results;

[0012] In the process of creating a decision tree, the optimal feature in the extracted features is selected as the node splitting basis of the decision tree, and the training data is used to recursively construct the decision tree until each leaf node contains samples belonging to the same class or meets the stopping criterion, then the decision tree construction is completed;

[0013] Step S4, obtaining a target feature reference set through a random forest;

[0014] Step S5, selecting the first N frames of the on-site detection picture as a training set, obtaining a suitable voting weight through continuous training, and then starting from the N+1 frame image, determining the final target through the determined weight and using the Adaboost algorithm for voting analysis.

[0015] In the step S1, the multi-stage filtering preprocessing includes the following steps, in particular:

[0016] Step S1.1, using an adaptive median filtering algorithm to filter the captured original image, in which process, the gray threshold is selected according to the maximum gray difference between the target and the background, so as to reduce the background noise and retain the target details;

[0017] Adaptive median filtering algorithm formula:

[0018]

[0019] Step S1.2, using a mean filter to further process the image processed in step S1.1, so as to further reduce the Gaussian noise in the image;

[0020] Step S1.3, using a 3x3 or 5x5 low-pass filter template to smooth the image processed in step S1.2 to remove irrelevant details in the image, blur the picture to get a rough description of the target to be captured;

[0021] Low-pass filter template:

[0022]

[0023] Step S1.4, by edge extension function, the target edge definition is enhanced.

[0024] In the step S2, when the shape feature of the image is extracted, the shape descriptor is used to describe the edge, width, length, area, principal axis, secondary axis, principal axis direction and aspect ratio of the target.

[0025] In the step S2, when the radiation feature of the image is extracted, the specific steps include:

[0026] Step S2.1.1, analyze the gray scale distribution of the target in the image, and extract the maximum gray scale, minimum gray scale and average gray scale of the target;

[0027] Step S2.1.2, construct the gray scale histogram of the target region, and determine the gray scale distribution mode;

[0028] Step S2.1.3, calculate the signal-to-noise ratio of the target region to measure the distinction between the target and the background.

[0029] In the step S2, when the motion feature of the image is extracted, the motion characteristics of the target in the time sequence image are evaluated by the time domain and spatial domain multi-scale differential features, including speed, acceleration, motion direction and its change;

[0030] The specific steps of the motion feature extraction include:

[0031] Step S2.2.1, track the position or pixel intensity of the target in the continuous image frame, and construct the time sequence data of the target;

[0032] Step S2.2.2, analyze the motion of the target at different time intervals or window sizes, differentiate the time sequence data, extract feature information in the differentiation result, the feature information includes average speed information, acceleration peak value information, frequency and amplitude information of speed change, and determine the motion characteristics of the target according to the above feature information.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] 1. Through deep learning and multi-feature analysis, it can effectively solve the problem of difficulty in distinguishing real targets from interference targets under the influence of external interference factors;

[0035] 2. By adopting the Adaboost algorithm and weighting the features to build an integrated classifier, the problem of being unable to accurately identify the real target in a multi-target scene with multiple suspected targets can be solved;

[0036] 3. By adopting the time domain and space domain feature change analysis, the problem of poor tracking stability for dynamic targets (especially targets that move quickly or change shape) can be solved;

[0037] 4. Therefore, the anti-interference target acquisition method based on machine vision technology measurement of the present invention can effectively solve the interference problem in the target capture process in complex scenes, and can effectively improve the accuracy and robustness of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention.

[0039] Figure 1 This is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be described below with reference to specific embodiments.

[0041] Example 1, as Figure 1 As shown, an anti-interference target acquisition method based on machine vision technology measurement is based on high-precision industrial cameras, computing servers and high-speed image acquisition cards. The high-precision industrial cameras are used to capture images of the scene in real time and ensure the clarity and continuity of the images. The computing servers are servers that meet the requirements of running complex image processing algorithms and machine learning models. The high-speed image acquisition cards are used to transmit the image data captured by the industrial cameras to the computing devices in real time.

[0042] Specifically, the anti-interference target acquisition method based on machine vision technology measurement includes the following steps:

[0043] Step S1, image preprocessing: performing multi-stage filtering preprocessing on the captured original image;

[0044] Step S2, target feature extraction: extracting features of the target in the image pre-processed in step S1, specifically including shape feature extraction, radiation feature extraction, and motion feature extraction;

[0045] In the shape feature extraction of the image, the shape descriptor is used to describe the edge, width, length, area, main axis, secondary axis, main axis direction and aspect ratio of the target;

[0046] Step S3, random forest construction: call the historical data sample library to obtain the anti-interference target acquisition library, randomly select M data points from the library, create a decision tree for the selected data points, each decision tree will generate a result, analyze the results comprehensively and output the majority vote or average;

[0047] In the process of creating a decision tree, the optimal feature in the extracted feature is selected as the node splitting basis of the decision tree, and then the training data is used to recursively construct the decision tree until each leaf node contains samples belonging to the same class or meets the stopping criterion, and then the decision tree construction is completed;

[0048] Step S4, obtain the target feature reference set through the random forest;

[0049] Step S5, select the first N frames of the on-site detection picture as the training set, obtain the appropriate voting weight by continuous training, and then start from the N+1 frame image, determine the final target by the determined weight and using the Adaboost algorithm for voting analysis;

[0050] Specifically, the steps of the Adaboost algorithm are:

[0051] a. First, assign the same weight to each training sample Where N is the number of samples;

[0052] b. Then for each round t = 1, 2,..., T:

[0053] 1) Train a weak classifier using weighted data;

[0054] 2) Calculate the error rate of the weak classifier Where I is the indicator function, yi is the actual label, and ht(xi) is the prediction of the classifier;

[0055] 3) Calculate the weight of the weak classifier

[0056] 4) Update the sample weight ωi←ωiexp(-αtyiht(xi));

[0057] c. Finally, construct the final classifier,

[0058] The anti-interference target acquisition method based on machine vision technology measurement of the embodiment one has the following advantages, specifically:

[0059] 1. Through deep learning and multi-feature analysis, the problem of distinguishing real targets from interference targets under the influence of external interference factors can be effectively solved;

[0060] 2. By using the Adaboost algorithm and weighting the features, an integrated classifier can be constructed to solve the problem of accurately identifying real targets in a multi-target scene with multiple suspected targets;

[0061] 3. By using time domain and spatial domain feature change analysis, the problem of poor tracking stability for dynamic targets, especially fast-moving or shape-changing targets, can be solved;

[0062] Based on the above, the anti-interference target acquisition method based on machine vision technology measurement of the first embodiment can effectively solve the interference problem in the target capture process in a complex scene, and can effectively improve the accuracy and robustness of target detection.

[0063] Embodiment two, the difference between embodiment two and embodiment one is that in step S1, the multi-stage filtering preprocessing includes the following steps, specifically:

[0064] Step S1.1, using an adaptive median filter algorithm to filter the captured original image, in this process, the gray threshold is selected according to the maximum gray difference between the target and the background, to reduce the background noise and retain the target details;

[0065] Adaptive median filter algorithm formula:

[0066]

[0067] Step S1.2, using a mean filter to further process the image processed by step S1.1, to further reduce the Gaussian noise in the image;

[0068] Step S1.3, using a 3x3 or 5x5 low-pass filter template to smooth the image processed by step S1.2, to remove irrelevant details in the image, and blur the picture to get a rough description of the target to be captured;

[0069] Low-pass filter template:

[0070]

[0071] Step S1.4, by using the edge extension function, the target edge definition is enhanced.

[0072] Embodiment three, the difference between embodiment three and embodiment one is that in step S2, when extracting the radiation features of the image, the specific steps include:

[0073] Step S2.1.1, analyzing the gray scale distribution of the target in the image, and extracting the maximum gray scale level, the minimum gray scale level, and the average gray scale level of the target;

[0074] Step S2.1.2, constructing a gray scale histogram of the target region, and determining the gray scale distribution mode;

[0075] Step S2.1.3, calculating the signal-to-noise ratio of the target region to measure the distinguishability of the target from the background.

[0076] Embodiment Four, the difference between this embodiment and Embodiment One is that, in the step S2, when performing motion feature extraction on the image, the motion characteristics of the target in the time series image are evaluated by using time domain and spatial domain multi-scale differential features, and the motion characteristics include speed, acceleration, motion direction, and changes thereof;

[0077] The specific steps of the motion feature extraction include:

[0078] Step S2.2.1, tracking the position or pixel intensity of the target in the continuous image frames, and constructing time series data of the target;

[0079] Step S2.2.2, analyzing the motion of the target at different time intervals or window sizes, differentiating the time series data, extracting feature information from the differentiation results, and determining the motion characteristics of the target according to the feature information, wherein the feature information includes average speed information, acceleration peak value information, frequency and amplitude information of speed change.

[0080] The above is only the preferred embodiment of the present application, and for those skilled in the art, the specific implementation manner and application range can be changed according to the idea of the present application, and the content of the description should not be understood as a limitation of the present application.

Claims

1. A method for obtaining an anti-interference target based on machine vision technology measurement, characterized in that: The following steps are included, specifically: Step S1, image preprocessing: performing multi-stage filtering preprocessing on the captured original image; Step S2, target feature extraction: extracting features of the target in the image pre-processed in step S1, specifically including shape feature extraction, radiation feature extraction, and motion feature extraction; When extracting motion features from images, the motion characteristics of the target in the time series images are evaluated through multi-scale differential features in the time domain and space domain. The motion characteristics include speed, acceleration, motion direction and their changes. When extracting radiation features from an image, the specific steps include: Step S2.1.1, analyzing the grayscale distribution of the target in the image, and extracting the maximum grayscale level, minimum grayscale level, and average grayscale level of the target; Step S2.1.2, constructing a grayscale histogram of the target area and determining the grayscale distribution pattern; Step S2.1.3, calculate the signal-to-noise ratio of the regional target to measure the discrimination between the target and the background; Step S3, random forest construction: call the historical data sample library to obtain the anti-interference target acquisition library, randomly select M data points from it, and create a decision tree for the selected data points. Each decision tree will produce a result, and the results will be comprehensively analyzed and output as a majority vote or average; In the process of creating a decision tree, the optimal feature among the extracted features is selected as the basis for node splitting of the decision tree, and then the decision tree is recursively constructed using the training data until the samples contained in each leaf node belong to the same category or the stopping criterion is met, and the decision tree is constructed. Step S4: Obtain a target feature reference set through random forest; Step S5: Select the first N frames of the on-site detection image as the training set, obtain appropriate counting weights through continuous training, and then start from the N+1th frame image, use the determined weights and the Adaboost algorithm to perform voting analysis to determine the final target.

2. The method for obtaining an anti-interference target based on machine vision technology measurement according to claim 1, characterized in that: In step S1, the multi-stage filtering preprocessing includes the following steps, specifically: Step S1.1: Use the adaptive median filter algorithm to filter the captured original image. During this process, the grayscale threshold is selected based on the maximum grayscale difference between the target and the background to reduce background noise and preserve target details: Adaptive median filtering algorithm formula: Step S1.2: further process the image processed in step S1.1 using a mean filter to further reduce Gaussian noise in the image; Step S1.3: Smoothing the image processed in step S1.2 using a 3×3 or 5×5 low-pass filter template to remove irrelevant details in the image and blur the image to obtain a rough description of the target to be captured; Low-pass filter template: Step S1.4: Enhance the target edge clarity through the edge extension function.

3. The method for obtaining an anti-interference target based on machine vision technology measurement according to claim 1, characterized in that: In step S2, when extracting shape features from an image, shape descriptors are used to describe the shape of the target, including its edge, width, length, area, major axis, minor axis, major axis direction, and aspect ratio.

4. The method for obtaining an anti-interference target based on machine vision technology measurement according to claim 1, characterized in that: In step S2, the specific steps of motion feature extraction include: Step S2.2.1, tracking the position or pixel intensity of the target in consecutive image frames to construct time series data of the target; Step S2.2.2: Analyze the target's motion at different time intervals or window sizes, differentiate the time series data, and extract feature information from the differentiation results. The feature information includes average velocity information, acceleration peak information, frequency and amplitude information of velocity changes, and determine the target's motion characteristics based on the above feature information.

Images