Target identification method

Abstract

The invention relates to the technical field of computers, and discloses a target recognition method, which comprises the steps of image preprocessing, feature extraction, feature selection, classifier training and classification, recognition result output, feedback and optimization, and can automatically learn deep features of a target through classification algorithms such as a convolutional neural network, thereby remarkably improving the recognition accuracy and improving the recognition efficiency. The misjudgment rate is reduced, a feature extraction method is continuously improved, and a higher-level feature descriptor and a feature matching algorithm are used, so that the recognition accuracy can be further improved; various types of images and videos including complex scenes such as different illumination conditions, angles and backgrounds can be processed, and the model can be quickly adapted and updated for a newly appearing target or the change of a target form so as to maintain the recognition accuracy; an image recognition algorithm is optimized, misjudged images serve as new training samples to be added into a training set again, more features and information are learned, and therefore the misjudgment rate in a future recognition task is reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, in particular to the field of image processing and recognition, and specifically to a target recognition method. Background Art

[0002] Object recognition is a computer vision technology that aims to identify objects in images or videos. When people look at photos or videos, they can easily recognize people, objects, scenes and visual details. The goal of object recognition is to teach computers to do what comes naturally to humans, that is, to gain a certain level of understanding of what the image contains.

[0003] After searching, the patent with application number CN202310127923.2 discloses a computer digital image target recognition model and method. The computer digital image target recognition model includes multiple groups of convolutional layers, multiple batch normalization layers and processing layer groups; each batch normalization layer corresponds to a group of convolutional layers, and the output nodes of the convolutional layers are connected to the input nodes of the corresponding batch normalization layers, the output nodes of the batch normalization layers are connected to the input nodes of the processing layer group, and the output nodes of the processing layer group are connected to the output layer; the convolutional layer is used to convolve the initial image to be recognized; the batch normalization layer is used to normalize the initial image, and the initial image is obtained based on the initial image to be recognized; the processing layer group is used to process the target image, and the target image is obtained based on the initial image; the output layer is used to output the recognition result, so that apple leaf diseases can be identified and the types of apple leaf diseases can be determined. This does not require manual recognition, can save manpower, and can ensure recognition accuracy.

[0004] The current computer-based target recognition method can only simply identify features in images, has poor image processing effects, cannot efficiently extract and select deep features in images, and reduces the recognition accuracy of targets in images. Therefore, we need to propose a target recognition method. Summary of the invention

[0005] The purpose of the present invention is to provide a target recognition method, which can automatically learn the deep features of the target through classification algorithms such as convolutional neural networks, thereby significantly improving the accuracy of recognition and reducing the misjudgment rate, and continuously improving the feature extraction method, using more advanced feature descriptors and feature matching algorithms, which can further improve the accuracy of recognition; it can process various types of images and videos, including complex scenes such as different lighting conditions, angles and backgrounds, and can quickly adapt and update the model to maintain the accuracy of recognition for newly emerging targets or changes in target morphology; and optimize the image recognition algorithm, re-add the misjudged images as new training samples to the training set, learn more features and information, thereby reducing the misjudgment rate in future recognition tasks, so as to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a target recognition method, comprising the following steps: S1. Image preprocessing: eliminate noise in the image, improve image clarity, adjust the pixel value of the image, and perform geometric transformation on the image; S2, feature extraction: extract useful feature information from the preprocessed image for subsequent classifier use. Feature information includes image edges, contours, and color values. S3, feature selection: based on feature extraction, further screen out features useful for target recognition, reduce the number of features, and improve the processing speed and accuracy of the classifier; S4. Classifier training and classification: Use positive and negative samples to train the classifier so that the classifier can learn to identify the target from the features extracted from the image; S5, recognition result output: output the recognition result according to the classification result of the classifier; S6. Feedback and optimization: During the target recognition process, the algorithm is continuously optimized through the feedback mechanism, and the images that the classifier misjudged are used as new training samples to retrain the classifier.

[0007] Preferably, in step S1, image preprocessing includes: Denoising: Use mean filtering to remove noise from the image, and calculate the mean of each pixel in the image and its surrounding pixels to smooth the image; Correct image pixels: Use histogram equalization to adjust the image's histogram to make the image's grayscale value evenly distributed and improve the image's contrast; use gamma correction to adjust the image's brightness to keep the image's visual effect good under different lighting conditions; use color correction to adjust the values ​​of the red, green, and blue channels to correct the image's color; Geometric transformation: Change the size of the image by scaling, correct the tilt or rotation of the target by rotation, move the image a certain distance horizontally or vertically by translation, and adjust the position of the target.

[0008] Preferably, in step S2, the process of feature extraction is as follows: S21, candidate object generation: setting a plurality of neighborhoods of different sizes and shapes in the image, and taking these pixel neighborhoods as candidate objects of salient regions; S22, visual saliency measurement: measure the visual saliency of the candidate object; S23, key area selection: based on the visual saliency measurement results, select the key areas in the image for subsequent processing.

[0009] Preferably, in step S3, the process of feature selection is as follows: S31, calculating the correlation index between each feature and the target variable; S32. Sort the features according to the correlation index, from high to low; S33, setting a correlation threshold, and selecting features above the threshold as retained features; S34. Consider the multicollinearity between features and eliminate redundant highly correlated features; S35. Record the features that are higher than the threshold and have not been eliminated as features useful for target recognition.

[0010] Preferably, in step S4, classifier training and classification include training the classifier and image classification, and the process of training the classifier is as follows: S41, training data and samples: positive samples are target images that need to be recognized, negative samples are target images that do not need to be recognized, and the number of positive samples and negative samples should be balanced; S42, secondary feature extraction: extract color, texture, and shape features from the image; S43, select a classification algorithm: select any one of the algorithms selected from convolutional neural network, support vector machine, decision tree, and random forest; S44, training: the secondary extracted features and the corresponding positive samples or negative samples are input into the classifier, the parameters of the classifier are adjusted through an iterative optimization algorithm to distinguish between the target and the background, and the performance of the classifier is evaluated using cross-validation.

[0011] Preferably, the process of image classification is as follows: S45, inputting an image to be identified: inputting the image to be identified into the trained classifier; S46, three-time feature extraction and classification: The classifier extracts features from the input image again and classifies the features using the parameters and rules learned during the training process. The classifier outputs a probability value, indicating the possibility that the image belongs to the target category; S47, decision and output: According to the output of the classifier, a threshold is set to determine whether the image contains the target. If the output probability value is higher than the threshold, the image is considered to contain the target; otherwise, the image is considered not to contain the target.

[0012] Preferably, in step S5, the output recognition result includes output content and output format, and the output content includes: Target type: The recognition result clearly indicates the target type contained in the image; Target location: The recognition result contains the location information of the target in the image, which is usually represented by a bounding box, that is, a rectangular box surrounding the target object, which is used to clearly indicate the specific area of ​​the target in the image; Confidence: A probability value that indicates the accuracy of the recognition result considered by the classifier.

[0013] Preferably, the output formats include: Text format: Output the target type, location and confidence information in text form for easy viewing and recording; Image format: Draw a bounding box and annotate the target type information on the original image, and output it in image format to help users quickly understand the recognition results.

[0014] Preferably, in step S6, the feedback mechanism includes false positive feedback. When the classifier makes a false positive, mistakenly identifying the target as the background or the background as the target, the false positive image is added back to the training set as a new training sample. The optimization content includes adjusting feature extraction and selection, optimizing the classifier algorithm, and data enhancement and preprocessing.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention can automatically learn the deep features of the target through classification algorithms such as convolutional neural networks, thereby significantly improving the accuracy of recognition and reducing the misjudgment rate. It also continuously improves the feature extraction method and uses more advanced feature descriptors and feature matching algorithms to further improve the accuracy of recognition. 2. The present invention can process various types of images and videos, including complex scenes with different lighting conditions, angles and backgrounds. It can quickly adapt and update the model to maintain recognition accuracy for new targets or changes in target morphology; 3. The present invention optimizes the image recognition algorithm, adds the misjudged images back into the training set as new training samples, learns more features and information, and thus reduces the misjudgment rate in future recognition tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the present invention; Figure 2 It is a flowchart of feature extraction of the present invention; Figure 3 A flowchart for selecting features of the present invention; Figure 4 It is a flowchart of the classifier training and classification process of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1-4The present invention provides a technical solution: a target recognition method, comprising the following steps: S1. Image preprocessing: eliminate noise in the image, improve image clarity, adjust the pixel value of the image, and perform geometric transformation on the image; In step S1, image preprocessing includes: Denoising: Use mean filtering to remove noise from the image, and calculate the mean of each pixel in the image and its surrounding pixels to smooth the image; Denoising is the first step in image preprocessing, and its purpose is to eliminate noise in the image caused by equipment problems, environmental interference, etc. Noise will reduce the clarity of the image and affect the subsequent feature extraction and recognition effects.

[0019] Correcting image pixels: Correcting image pixels is an important step to ensure the consistency of image color and brightness. This includes adjusting the image's contrast, brightness, and color balance to make the image more consistent with the requirements of subsequent processing.

[0020] Histogram equalization is used to adjust the histogram of the image to make the grayscale value of the image evenly distributed and improve the contrast of the image; gamma correction is used to adjust the brightness of the image to keep the image in good visual effect under different lighting conditions; color correction is used to adjust the values ​​of the red, green and blue channels to correct the color of the image; Geometric transformation: Geometric transformation is to scale, rotate, translate and other operations on the image to adapt to objects of different sizes and angles. These transformations can ensure that the position and posture of the object in the image meet the requirements of subsequent processing.

[0021] Scaling changes the size of the image, rotation corrects the tilt or rotation of the target, and translation moves the image a certain distance horizontally or vertically to adjust the position of the target.

[0022] Image preprocessing also includes normalization: Normalization is the process of converting image data into a unified standard form for subsequent processing. Normalization can eliminate data differences between different images and improve the robustness and stability of the algorithm.

[0023] Grayscale normalization: scale the grayscale value of the image to a fixed range (such as 0-1 or 0-255); Histogram normalization: make the histogram of the image conform to a certain distribution, such as normal distribution or uniform distribution; Z-score normalization: Normalize the data according to its mean and standard deviation so that the data has zero mean and unit variance.

[0024] S2, feature extraction: extract useful feature information from the preprocessed image for subsequent classifier use. Feature information includes image edges, contours, and color values. In step S2, the feature extraction process is as follows: S21, candidate object generation: setting a plurality of neighborhoods of different sizes and shapes in the image, and taking these pixel neighborhoods as candidate objects of salient regions; S22. Visual saliency measurement: measures the visual saliency of candidate objects; it is the core link in the image region detection process. Salient features can be extracted through internal extraction methods (such as symmetry, complexity, inconsistency, etc. based on pixel neighborhood) or external extraction methods (such as novel stimuli formed by contrast with the outside world).

[0025] S23, key area selection: based on the visual saliency measurement results, select the key areas in the image for subsequent processing.

[0026] S3, feature selection: based on feature extraction, further screen out features useful for target recognition, reduce the number of features, and improve the processing speed and accuracy of the classifier; In step S3, the purpose of feature selection is to reduce the number of features, improve model performance, and enhance model interpretability; Reduce the number of features: By removing redundant or irrelevant features, the dimension of the data is reduced, thereby reducing the computational cost and speeding up the training process; Improve model performance: Removing irrelevant features can reduce noise and make the model more focused on key information, thereby improving the accuracy of predictions; Improve model interpretability: Simplify the model structure to help understand the intrinsic structure of the data and improve the interpretability of the model.

[0027] The process of feature selection is as follows (filtering method): S31, calculating the correlation index between each feature and the target variable; S32. Sort the features according to the correlation index, from high to low; S33, setting a correlation threshold, and selecting features above the threshold as retained features; S34. Consider the multicollinearity between features and eliminate redundant highly correlated features; S35. Record the features that are higher than the threshold and have not been eliminated as features useful for target recognition.

[0028] Optionally, feature selection also includes a wrapper method: Choose a basic machine learning model, such as random forest, logistic regression, etc.

[0029] Adopt heuristic feature subset search strategies, such as forward selection, backward elimination, etc.

[0030] For each candidate feature subset, cross-validation and other methods are used to evaluate the performance indicators of the model on this feature subset, such as accuracy, F1 value, etc.

[0031] According to the results of the performance indicators, the optimal feature subset is selected.

[0032] The selection of feature subsets can be optimized using methods such as Grid Search or RandomSearch.

[0033] The final selected feature subset is the result of the wrapping method.

[0034] Optionally, feature selection also includes embedding methods: Choose a machine learning model with a built-in feature selection mechanism, such as LASSO regression, Ridge regression, decision tree, etc.

[0035] While training these models, the models automatically learn the importance of features and perform feature selection.

[0036] LASSO will generate sparse feature weight vectors, thereby achieving feature selection.

[0037] The decision tree model selects the optimal splitting feature based on criteria such as information gain or Gini coefficient.

[0038] The final output feature importance or weight vector is the result of the embedding method selection.

[0039] S4. Classifier training and classification: Use positive and negative samples to train the classifier so that the classifier can learn to identify the target from the features extracted from the image; In step S4, classifier training and classification include training the classifier and image classification. The process of training the classifier is as follows: S41, training data and samples: positive samples are target images that need to be recognized, negative samples are target images that do not need to be recognized, and the number of positive samples and negative samples should be balanced; S42, secondary feature extraction: extract color, texture, and shape features from the image; S43, select a classification algorithm: select any one of the algorithms selected from convolutional neural network, support vector machine, decision tree, and random forest; S44, training: the secondary extracted features and the corresponding positive samples or negative samples are input into the classifier, the parameters of the classifier are adjusted through an iterative optimization algorithm to distinguish between the target and the background, and the performance of the classifier is evaluated using cross-validation.

[0040] The process of image classification is as follows: S45, inputting an image to be identified: inputting the image to be identified into the trained classifier; S46, three-time feature extraction and classification: The classifier extracts features from the input image again and classifies the features using the parameters and rules learned during the training process. The classifier outputs a probability value, indicating the possibility that the image belongs to the target category; S47, decision and output: According to the output of the classifier, a threshold is set to determine whether the image contains the target. If the output probability value is higher than the threshold, the image is considered to contain the target; otherwise, the image is considered not to contain the target.

[0041] S5, recognition result output: output the recognition result according to the classification result of the classifier; In step S5, the output recognition result includes output content and output format, and the output content includes: Target type: The recognition result clearly indicates the type of target contained in the image; for example, in a face recognition task, it will indicate to which individual or group of people the recognized face belongs; in a vehicle recognition task, it will indicate the type of vehicle (such as a car, truck, motorcycle, etc.).

[0042] Target location: The recognition result contains the location information of the target in the image, which is usually represented by a bounding box, that is, a rectangular box surrounding the target object, which is used to clearly indicate the specific area of ​​the target in the image; Confidence: A probability value that indicates the accuracy of the recognition result considered by the classifier.

[0043] Output formats include: Text format: Output the target type, location and confidence information in text form for easy viewing and recording; Image format: Draw a bounding box and annotate the target type information on the original image, and output it in image format to help users quickly understand the recognition results.

[0044] S6. Feedback and optimization: During the target recognition process, the algorithm is continuously optimized through the feedback mechanism, and the images that the classifier misjudged are used as new training samples to retrain the classifier.

[0045] In step S6, the feedback mechanism includes misjudgment feedback. When the classifier misjudges, mistakenly identifying the target as background or the background as the target, the misjudged image is added back to the training set as a new training sample. The feedback mechanism also includes: User feedback: Users may provide feedback on the recognition results. For example, users may think that a certain recognition result is inaccurate or the position of the bounding box is inaccurate. User feedback can be used as additional training data or as a basis for adjusting parameters to help optimize the performance of the classifier and the accuracy of the output results; Automatic evaluation and feedback: Establish an automatic evaluation system to continuously monitor and evaluate the performance of the classifier. For example, you can calculate the accuracy, recall, F1 score and other indicators of the classifier, and judge whether the performance of the classifier is stable or declining based on the changes in these indicators. When the evaluation indicator is lower than the preset threshold, the feedback mechanism can be triggered to automatically collect new training samples or adjust the parameters of the classifier to improve its performance.

[0046] The optimization content includes adjusting feature extraction and selection, optimizing classifier algorithms, data enhancement and preprocessing.

[0047] Adjust feature extraction and selection: Try using different feature extraction methods (such as convolutional neural networks in deep learning, SIFT, SURF, etc. in traditional methods) and feature selection algorithms (such as filtering, wrapping, embedded, etc.) to find the feature set that best suits the current task; by adjusting the parameters and methods of feature extraction and selection, the input data of the classifier can be optimized, thereby improving its recognition performance.

[0048] Optimize the classifier algorithm: Use different classification algorithms (such as support vector machines, decision trees, random forests, neural networks, etc.) and select them according to the actual task requirements and data characteristics. For the selected classification algorithm, its performance can be optimized by adjusting its parameters (such as learning rate, regularization term, number of iterations, etc.).

[0049] Data augmentation and preprocessing: Data augmentation is a method of generating more training samples by transforming the original data. For example, images can be rotated, scaled, flipped, etc. to generate more training samples and increase the diversity of the data; Data preprocessing is to clean, normalize, and standardize the raw data to improve the quality and consistency of the data. Through data enhancement and preprocessing, the input data of the classifier can be optimized, thereby improving its recognition performance.

[0050] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A target recognition method, characterized in that: The following steps are involved: S1. Image preprocessing: eliminate noise in the image, improve image clarity, adjust the pixel value of the image, and perform geometric transformation on the image; S2, feature extraction: extract useful feature information from the preprocessed image for subsequent classifier use. Feature information includes image edges, contours, and color values. S3, feature selection: based on feature extraction, further screen out features useful for target recognition, reduce the number of features, and improve the processing speed and accuracy of the classifier; S4. Classifier training and classification: Use positive and negative samples to train the classifier so that the classifier can learn to identify the target from the features extracted from the image; S5, recognition result output: output the recognition result according to the classification result of the classifier; S6. Feedback and optimization: During the target recognition process, the algorithm is continuously optimized through the feedback mechanism, and the images that the classifier misjudged are used as new training samples to retrain the classifier.

2. The target recognition method according to claim 1, characterized in that: In step S1, image preprocessing includes: Denoising: Use mean filtering to remove noise from the image, and calculate the mean of each pixel in the image and its surrounding pixels to smooth the image; Correct image pixels: Use histogram equalization to adjust the image's histogram to make the image's grayscale value evenly distributed and improve the image's contrast; use gamma correction to adjust the image's brightness to keep the image's visual effect good under different lighting conditions; use color correction to adjust the values ​​of the red, green, and blue channels to correct the image's color; Geometric transformation: Change the size of the image by scaling, correct the tilt or rotation of the target by rotation, move the image a certain distance horizontally or vertically by translation, and adjust the position of the target.

3. The target recognition method according to claim 1, characterized in that: In step S2, the feature extraction process is as follows: S21, candidate object generation: setting a plurality of neighborhoods of different sizes and shapes in the image, and taking these pixel neighborhoods as candidate objects of salient regions; S22, visual saliency measurement: measure the visual saliency of the candidate object; S23, key area selection: based on the visual saliency measurement results, select the key areas in the image for subsequent processing.

4. The target recognition method according to claim 1, characterized in that: In step S3, the process of feature selection is as follows: S31, calculating the correlation index between each feature and the target variable; S32. Sort the features according to the correlation index, from high to low; S33, setting a correlation threshold, and selecting features above the threshold as retained features; S34. Consider the multicollinearity between features and eliminate redundant highly correlated features; S35. Record the features that are higher than the threshold and have not been eliminated as features useful for target recognition.

5. The target recognition method according to claim 1, characterized in that: In step S4, classifier training and classification include training the classifier and image classification. The process of training the classifier is as follows: S41, training data and samples: positive samples are target images that need to be recognized, negative samples are target images that do not need to be recognized, and the number of positive samples and negative samples should be balanced; S42, secondary feature extraction: extract color, texture, and shape features from the image; S43, select a classification algorithm: select any one of the algorithms selected from convolutional neural network, support vector machine, decision tree, and random forest; S44, training: the secondary extracted features and the corresponding positive samples or negative samples are input into the classifier, the parameters of the classifier are adjusted through an iterative optimization algorithm to distinguish between the target and the background, and the performance of the classifier is evaluated using cross-validation.

6. The target recognition method according to claim 5, characterized in that: The process of image classification is as follows: S45, inputting an image to be identified: inputting the image to be identified into the trained classifier; S46, three-time feature extraction and classification: The classifier extracts features from the input image again and classifies the features using the parameters and rules learned during the training process. The classifier outputs a probability value, indicating the possibility that the image belongs to the target category; S47, decision and output: According to the output of the classifier, a threshold is set to determine whether the image contains the target. If the output probability value is higher than the threshold, the image is considered to contain the target; otherwise, the image is considered not to contain the target.

7. The target recognition method according to claim 1, characterized in that: In step S5, the output recognition result includes output content and output format, and the output content includes: Target type: The recognition result clearly indicates the target type contained in the image; Target location: The recognition result contains the location information of the target in the image, which is usually represented by a bounding box, that is, a rectangular box surrounding the target object, which is used to clearly indicate the specific area of ​​the target in the image; Confidence: A probability value that indicates the accuracy of the recognition result considered by the classifier.

8. The target recognition method according to claim 7, characterized in that: Output formats include: Text format: Output the target type, location and confidence information in text form for easy viewing and recording; Image format: Draw a bounding box and annotate the target type information on the original image, and output it in image format to help users quickly understand the recognition results.

9. The target recognition method according to claim 1, characterized in that: In step S6, the feedback mechanism includes false positive feedback. When the classifier makes a false positive, mistakenly identifying the target as the background or the background as the target, the false positive image is added back to the training set as a new training sample. The optimization content includes adjusting feature extraction and selection, optimizing the classifier algorithm, and data enhancement and preprocessing.

Images