A classification method for distinguishing highly similar target objects

Abstract

The present invention provides a classification method for distinguishing target objects with high similarity, including: defining a set of high-similarity objects, including error-prone categories, associated categories, etc.; obtaining environmental images of different positions during normal driving of an unmanned vehicle on a highway; detecting multiple objects in the environmental images using target detection technology; screening object label categories belonging to the set of high-similarity objects; training a twin neural network using the idea of ​​meta-learning to enable it to have the ability to judge whether the pixel ratio of the main body in the image has changed; inputting before-and-after comparison photos of the same object in environmental images at different positions into the twin neural network to obtain classification results; based on the perspective projection effect that the pixel ratio change rate of objects at different distances is different, objects with a smaller change rate are distant objects, and objects with a larger change rate are nearby, the object pixel ratio change rate is mapped to the distance factor to accurately distinguish objects with high similarity, thereby effectively improving the classification effect.

Description

Technical Field

[0001] The present invention relates to the field of image classification in a target detection framework, and in particular to a classification method for distinguishing target objects with high similarity. Background Art

[0002] Object detection is a crucial task in computer vision, aiming to identify the location and category of a target object in an image or video. Object detection is a crucial task in computer vision, widely used in intelligent transportation, security monitoring, autonomous driving, facial recognition, and other fields. Object detection can generally be divided into two stages: object extraction and object classification. Object extraction typically uses techniques such as edge detection, image segmentation, or feature extraction to separate the target from the background in an image. Object classification, on the other hand, categorizes the extracted targets into different categories and is typically implemented using deep learning algorithms such as recurrent neural networks and convolutional neural networks.

[0003] In recent years, deep learning methods, particularly convolutional neural networks, have demonstrated impressive performance in object classification tasks. The basic idea is to extract image features through multiple convolutional and pooling layers, and then perform classification through fully connected layers. Trained on large-scale datasets, deep learning methods can automatically learn deep image features, resulting in improved performance and generalization in object classification tasks. Object classification is also a fundamental task in computer vision, aiming to classify objects in an input image or video into predefined categories. However, when objects are highly similar, classification models can misclassify them.

[0004] In a real-world scenario, misclassification caused by highly similar objects occurred in Tesla's Full Self-Driving (FSD) system. The system mistakenly identified the moon as a yellow light, causing the vehicle to slow down while cruising at high speed. This was due to the high feature similarity between the moon and the yellow light, preventing the convolutional neural network from correctly distinguishing between them. Similar misclassification issues exist not only in autonomous vehicles but also in other fields. For example, unmanned combat aircraft also face difficulties identifying similar objects during visual navigation or when operating in jungles or near low-lying buildings. The root cause of these misclassification issues is the small numerical differences between the objects in their feature vectors, making it difficult for the model to distinguish them. However, current convolutional neural networks still cannot avoid misclassification when dealing with highly similar objects, which can lead to incorrect decisions and negative consequences. As in the above example, model misclassification can lead to serious traffic accidents and significant losses. Therefore, improving and resolving the problem of neural network misclassification of highly similar objects is urgent. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention introduces a meta-learning mechanism to overcome the difficulties of small sample training, and uses non-visual factors to achieve the purpose of accurate identification of highly similar targets; it provides a classification method for distinguishing highly similar target objects to solve the existing problem of misclassification in dealing with highly similar target scenarios, and effectively improves the classification accuracy of neural networks.

[0006] A classification method for distinguishing highly similar target objects, specifically including the following contents:

[0007] Step 1: Define a set of high similarity categories;

[0008] The high similarity categories are specifically: pairs or groups of categories with similar features or attributes in the dataset, which are highly similar in color, shape and texture;

[0009] The high-similarity category set includes error-prone categories, identical categories, and similar categories;

[0010] The error-prone category: objects that are very similar to other objects in appearance, shape, or function and are easy to be confused or mistaken;

[0011] The same category: multiple objects belong to the same category, and objects in this category have the same characteristics or functions;

[0012] Similar categories: objects in the same category have appearance features of similar shape, texture or color;

[0013] Step 2: Use the YOLO v3 target detection framework to screen and identify objects with high similarity categories;

[0014] The device uses its onboard sensors to capture environmental images at different locations during its motion. These environmental source images at different device locations are then fed into the YOLO v3 object detection framework. The YOLO v3 object detection framework outputs the location, category, and confidence level of the detected target object. Each target object detection result is screened, requiring its confidence level to be above a preset threshold and the target object category to belong to a set of highly similar categories, thereby filtering out easily confused target information.

[0015] The device is a terminal and intelligent system with target detection requirements; the sensor carried by the device is an instrument for capturing images and videos, and is used to implement target detection applications;

[0016] The environmental images at different positions during the movement are acquired, wherein the intervals between the positions of the environmental images are dynamically adjusted, and the length of the intervals is determined by the target detection output result of the environmental source image acquired at the previous position; if the environmental source image detects that an object belongs to a high-similarity object set, the frequency of environmental acquisition is increased; otherwise, the frequency of environmental acquisition is reduced;

[0017] Step 3: Group pre-training Siamese twin neural network;

[0018] The Siamese twin neural network consists of two identical sub-networks, each with the same structure and weights, which are used to process paired inputs and learn the similarities and differences between paired images;

[0019] The idea of ​​meta-learning is used to train the Siamese twin neural network, enabling it to accurately determine whether the pixel ratio of the main part in two images has changed; the steps are as follows:

[0020] Step 3.1: Dataset preprocessing: Based on the ImageNet dataset, the dataset was divided into a training set and a test set in an 8:2 ratio. The two datasets were divided into two groups for training and testing respectively. One group had the pixel ratio of the main part of the image changed and was labeled 1; the other group had the same pixel ratio and was labeled 0. The pixel ratio is used to describe the ratio of the number of pixels of the object in the image to the total number of pixels in the image, which is the proportion of the object to the total image area.

[0021] Step 3.2: Construct a Siamese twin neural network: Use a convolutional neural network as the basic structure of the Siamese twin neural network; the mathematical model of the Siamese twin network is as follows:

[0022]

[0023] Where θ is the model parameter, N is the number of samples in the data set, and L i () represents the loss function of the i-th pair of samples, and Represent the sample sets used for training and testing respectively;

[0024] Step 3.3: Define the loss function: Based on the binary classification problem, use the binary cross entropy loss function to measure the difference between the Siamese twin neural network prediction results and the true label; the loss function expression is as follows:

[0025]

[0026] Where N is the number of samples in the dataset, y i is the true label of the i-th pair of samples, p i is the label predicted by the Siamese twin network;

[0027] The loss function measures the difference between the Siamese twin network's predictions for the input images and their true labels. When the twin network's predictions for a pair of input images are the same as their true labels, the loss function value is small. Conversely, when the predictions are different from the true labels, the loss function value is large. By minimizing the loss function, the Siamese twin network will gradually learn how to determine whether the pixel ratios of objects in the two images have changed, and achieve more accurate prediction results by adjusting parameters.

[0028] Step 3.4: Follow the above steps to pre-train the Siamese twin network in groups: use the paired images as the input of the Siamese twin network, calculate the loss function, and finally use the backpropagation algorithm to update the network parameters;

[0029] Step 4: Use the Siamese twin network trained in step 3 to perform target classification and output the label classification results;

[0030] The trained Siamese twin network is used for the target object classification task of the target detection framework. By screening the environmental source images at different locations in step 2 to obtain images with the same easily confused object information, images of the same object at different locations are intercepted and input into the Siamese twin network in pairs, and the classification results are output. If the pixel ratio of the main part in the two images changes significantly, the output label is 1; if the pixel ratio of the main part changes slightly, the output label is 0.

[0031] Step 5: Distinguish objects with high similarity based on label classification results;

[0032] Step 5.1: Obtain the Siamese twin network classification result label: If the classification result label is 1, it means that the object pixel ratio has changed significantly; otherwise, the label is 0, which means that the object pixel ratio has not changed significantly.

[0033] Step 5.2: Based on the perspective projection effect, the closer an object is to the observer in three-dimensional space, the larger its projection is in the two-dimensional image, and the farther the object is in three-dimensional space, the smaller its projection is in the two-dimensional image. If the pixel ratio of an object in images at different positions does not change, it means that the object is farther away from the device sensor. Conversely, a significant change in the pixel ratio indicates that the object is closer to the device sensor. Successfully distinguish highly similar objects based on the non-visual attribute of distance.

[0034] The present invention has the following beneficial technical effects:

[0035] Distinguishing highly similar objects is an important research direction in the field of computer vision. This invention improves image classification accuracy and generalization capabilities, and can successfully distinguish highly similar objects in fields such as unmanned driving, medical diagnosis, and industrial production, thereby improving safety and production efficiency and promoting progress and development in the field of computer vision. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a classification method for distinguishing high-similarity targets according to the present invention.

[0037] Figure 2 A real-world example of misclassification in self-driving cars in real-world applications.

[0038] Figure 3 This is the framework diagram of the pre-trained Siamese twin neural network of the present invention.

[0039] Figure 4 This is the detection result diagram of the target detection framework simulating an unmanned vehicle in the present invention, and the scene image is taken at a distance from the traffic light.

[0040] Figure 5 This is the detection result diagram of the target detection framework simulating an unmanned vehicle in the present invention, and the scene image is taken at a position close to the traffic light.

[0041] Figure 6 This is an example of the group training set used to pre-train the Siamese twin network in the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments;

[0043] The current mainstream classification network is not enough to distinguish objects with highly similar features and highly consistent appearance; the application example of this invention is the post-classification optimization part of the target detection framework of unmanned vehicles. Figure 2 The neural network equipped in Tesla's self-driving car mistakenly identified the moon as a yellow light, which affected the driving logic of the entire vehicle and caused it to repeatedly tap the brakes. This misclassification may cause serious traffic accidents in the field of self-driving cars, which in turn seriously threatens people's lives and property. To this end, the present invention provides a classification method for distinguishing high-similarity targets, which can be applied to scenarios such as traffic identification and medical diagnosis. By further combining meta-learning ideas with non-visual factors, it can achieve accurate classification of high-similarity objects, effectively solving the misclassification problem of high-similarity objects in complex scenarios, thereby improving the traffic safety of self-driving cars.

[0044] A classification method to distinguish high similarity targets, such as Figure 1 As shown, specifically including the following:

[0045] Step 1: Define a set of high similarity categories;

[0046] Highly similar categories are specifically pairs or groups of categories with similar features or attributes in the dataset, which are highly similar in color, shape, and texture. Because the feature differences between highly similar categories are small, ordinary neural networks cannot accurately distinguish them.

[0047] The highly similar category set includes error-prone categories, identical categories, and similar categories, as well as categories that are difficult for standard image classification networks and modules to distinguish correctly or are easily confused. The highly similar object set refers to sample categories with similar features or attributes in the dataset. These categories are often highly similar in appearance, features, and attributes, making them difficult even for the human eye to distinguish. In classification problems, highly similar categories can make it difficult for image classifiers to accurately distinguish samples from different categories, thus affecting classifier performance.

[0048] The error-prone category: objects that are very similar to other objects in appearance, shape, or function and are easy to be confused or mistaken;

[0049] The same category refers to multiple objects belonging to the same category. Objects in the same category have the same characteristics or functions. For example, all dogs belong to the Canidae family, and there are only subtle differences between different types of dogs. In this case, it is difficult for the neural network to detect the subtle differences between different types of dogs, resulting in misclassification.

[0050] Similar categories: objects in the same category have similar appearance features such as shape, texture or color; for example, all round objects and objects with similar texture features.

[0051] In the implementation of the present invention, the highly similar object set S is defined as: S = {(moon, yellow light), (sun, red light), (cloud sticker, real clouds in the sky)}. The implementation of the present invention assumes that the elements in the highly similar object set S typically exist in pairs or groups, and the categories of each group of elements in the set have a high degree of similarity to each other. For example, the moon and yellow light have a high degree of similarity, and ordinary classification neural networks have difficulty accurately distinguishing between these two categories. The highly similar object set S listed in the present invention includes but is not limited to the three groups of element categories listed above, and can be supplemented and modified according to actual needs.

[0052] Step 2: Use the YOLO v3 object detection framework to identify objects of high similarity categories;

[0053] The device's onboard sensors capture environmental images at different locations during its motion, and these environmental source images at different device locations are fed into the YOLO v3 object detection framework. The YOLO v3 object detection framework outputs the location, category, and confidence level of the detected target object. Each target object detection result is screened, requiring its confidence level to be above a preset threshold and the target object category to belong to a set of highly similar categories, thereby obtaining information about easily confused targets.

[0054] The device is a terminal and intelligent system with target detection requirements; the sensor carried by the device is an instrument for capturing images and videos, and is used to implement target detection applications;

[0055] The environmental images at different positions during the motion are acquired, wherein the intervals between the positions of the environmental images are dynamically adjusted, and the length of the intervals is determined by the target detection output result of the environmental source image acquired at the previous position; if the environmental source image detects that an object belongs to a high-similarity object set, the frequency of environmental acquisition is increased to timely correct the object classification result; otherwise, the frequency of environmental acquisition is reduced to save computing costs;

[0056] In the embodiment of the present invention, images of different positions of an unmanned vehicle in a normal driving state on a highway are obtained, such as Figure 4 、 5 As shown in the figure, a self-driving car captures the first environmental source image P1 while driving normally on a highway. After traveling a distance d, it captures the second image P2. The distance d is set based on whether the target detection output result of the previous recognition of the captured environmental source image P1 contains a category in the similar category set S. If the detection result of the previous recognition of image P1 contains a category in the highly similar category set S, the driving distance d is appropriately reduced, and the recognition frequency is increased within a shorter distance interval to quickly calibrate the classification detection results.

[0057] This embodiment of the present invention uses YOLO v3 to simulate the target detection framework used in real self-driving cars, addressing the misclassification problem often encountered in the classification module of conventional target detection frameworks. YOLO v3 is used to simulate the detection output of the target detection framework. The captured environmental source images P1 and P2 are sequentially input into the target detection framework in the order in which they were captured. The output is the location information and type information O = {bbox, score, label} for each target object in the environmental image. Bbox = (x1, y1, x2, y2) contains the object's boundary information, and score and label represent the classification confidence and object classification category, respectively.

[0058] Set the condition to filter the object label in S that belongs to the above similar category set, and require the corresponding confidence to be greater than the rated threshold score>value(0.3). Here the rated threshold is 0.3 to ensure that the corresponding classification category has a high confidence. After conditionally filtering the target detection output results of the environment source images P1 and P2, pairs of objects with high similarity can be obtained, such as: 11 , O 12 and O 21 , O 22 .

[0059] Step 3: Group pre-training Siamese twin neural network; as shown in the attached Figure 3 As shown;

[0060] The Siamese twin neural network consists of two identical sub-networks, each with the same structure and weights, which are used to process paired inputs and learn the similarities and differences between paired images;

[0061] The idea of ​​meta-learning is used to train the Siamese twin neural network, enabling it to accurately determine whether the pixel ratio of the main part in two images has changed; the steps are as follows:

[0062] Step 3.1: Dataset preprocessing: The ImageNet Nips2017 competition dataset is used, which contains 1000 image samples of different categories. The dataset is divided into a training set and a test set in a ratio of 8:2, with 800 image samples used for the training set and 200 image samples used for the test set. The grouped training set cited in the embodiment of the present invention is shown in the attached figure. Figure 6 As shown in Figure 1, the YOLO v3 framework is used to extract the image region containing the main object from the training data. The entire training dataset is divided into two groups, each containing 400 images. In the first group, the main object of the image is stretched and transformed, and these images are labeled as In the second set, the images remain unchanged and are labeled (x, x, 0). In these paired training data, label 1 indicates that one of the images in the pair has had its main object stretched, changing its pixel ratio in the image. Label 0 indicates that the main object in the image has not been transformed. The pixel ratio is used to describe the ratio of the number of pixels of the object in the image to the total number of pixels in the image, which is the ratio of the object to the total image area.

[0063] Step 3.2: Construct a Siamese twin neural network: Use a convolutional neural network as the basic structure of the Siamese twin neural network; the mathematical model of the Siamese twin network is as follows:

[0064]

[0065] Where θ is the model parameter, N is the number of samples in the data set, and L i () represents the loss function of the i-th pair of samples, and Represent the sample sets used for training and testing respectively;

[0066] Step 3.3: Define the loss function: Based on the binary classification problem, use the binary cross entropy loss function to measure the difference between the Siamese twin neural network prediction results and the true label; the loss function expression is as follows:

[0067]

[0068] Where N is the number of samples in the dataset, y i is the true label of the i-th pair of samples, p i is the label predicted by the Siamese twin network;

[0069] The loss function measures the difference between the Siamese twin network's predictions for the input images and their true labels. When the twin network's predictions for a pair of input images are the same as their true labels, the loss function value is small. Conversely, when the predictions are different from the true labels, the loss function value is large. By minimizing the loss function, the Siamese twin network will gradually learn how to determine whether the pixel ratios of objects in the two images have changed, and achieve more accurate prediction results by adjusting parameters.

[0070] Step 3.4: Follow the above steps to pre-train the Siamese twin network in groups: use the paired images as input to the Siamese twin network, calculate the loss function, and finally use the backpropagation algorithm to update the network parameters. The Siamese twin network uses a binary classifier to determine whether the pixel ratio of the main part of the two images has changed. The output of the binary classifier is represented as y = {0, 1}, where y = 0 indicates that the pixel ratio of the main part in the two images has not changed, and y = 1 indicates that it has changed.

[0071] After pre-training, the Siamese twin network's performance was evaluated on a test set consisting of 200 pairs of image samples. Using the same method as the training set, the main subject in 100 images was stretched and transformed, changing the pixel ratio of the main object. This test demonstrated the Siamese twin network's ability to distinguish between changes in the pixel ratio of the main object between the two images. The pre-trained Siamese twin network was used to make predictions on the processed test set, and the average accuracy was calculated to evaluate the network's performance. The final experimental results showed that the pre-trained Siamese twin network achieved 100% accuracy on the test set. This demonstrates that the Siamese twin network has a strong ability to distinguish between changes in the pixel ratio of the main object in the image samples.

[0072] Step 4: Use the Siamese twin network trained in step 3 to perform target classification and output the label classification results;

[0073] As described in the third step of group pre-training Siamese twin networks, the meta-learning-based Siamese twin network has strong generalization ability and adaptability, and can quickly adapt to new unseen tasks.

[0074] The trained Siamese twin network is used for the target object classification task in the target detection framework. As described in step 2, the object detection results of the target detection framework are filtered to obtain easily confused target information. Images of the same easily confused object information are obtained from environmental source images at different locations. These images are input into the Siamese twin network as paired images, and the classification results are output. If the pixel ratio of the main part of the two images changes significantly, the output label is 1; if the pixel ratio of the main part changes slightly, the output label is 0.

[0075] In the implementation of the present invention, the environmental source images taken by the unmanned vehicle at different distances and positions are input into the YOLOv3 target detection framework respectively, and a group of objects of the same easily confused object at different distances and positions are obtained after the output result conditions of the target detection framework are filtered. 11 and O 12. The object is represented as O = {bbbox, label, score}. Compare the bounding boxes in the two images containing the same object, and select the bounding box with the larger pixel area as the standard. Then stretch and expand the smaller bounding box in the other image so that its range matches the standard bounding box. In this way, two images containing the same object are obtained, but the pixel area ratio of the object in the image is different. The two processed images are input into the Siamese twin network to construct the support set (support set) and query set (query set). The Siamese twin network outputs the prediction result to determine whether the two images have obvious pixel ratio changes. If the twin network output result is 1, the pixel ratio of the main part of the two images has changed significantly; if the twin network output result is 0, the pixel ratio of the main part of the two images has changed slightly.

[0076] Step 5: Distinguish objects with high similarity based on the label classification results;

[0077] Conventional classification neural networks struggle to distinguish highly similar objects. This is primarily because, from an image feature perspective, there is a significant degree of overlap between highly similar objects. Therefore, this study introduces distance, a non-visual attribute, to help effectively distinguish highly similar objects. The specific steps are as follows:

[0078] Step 5.1: Obtain the Siamese twin network classification result label: As described in step 4, input the images of the same easily confused object at different locations into the Siamese twin network in pairs; if the Siamese twin network outputs a classification result label of 1, it means that the object pixel ratio has changed significantly, otherwise the label is 0, which means that the object pixel ratio has not changed significantly;

[0079] Step 5.2: According to the perspective projection effect, the closer the object in the three-dimensional space is to the observer, the larger its projection in the two-dimensional image is, and the farther the object in the three-dimensional space is, the smaller its projection in the two-dimensional image is; if the pixel ratio of the easily confused object in the image at different positions does not change, it means that the easily confused object is far away from the device sensor; on the contrary, if the pixel ratio changes significantly, it means that the easily confused object is close to the device sensor. In the implementation of the present invention, if Figure 2As shown, in the unmanned vehicle environment, the classification module of the on-board target detection framework mistakenly identifies the moon as a yellow light. From the perspective of image features, the appearance features such as color and shape of the moon and the yellow light category are very similar. Therefore, the present invention introduces the non-visual attribute of distance to distinguish high-similarity objects. In this example, if the unmanned vehicle recognizes that the yellow light category belongs to the high-similarity category set, it will quickly shorten the detection interval d, and capture the same yellow light object image at different positions and input it into the Siamese twin network at the same time. If the twin network outputs a label result of 0, it means that the pixel ratio of this target object has changed little and is far away from the current car. At this time, the current high-similarity object is the moon category rather than the yellow light category. The present invention can verify the high-similarity object category in a timely manner, effectively avoiding misclassification.

[0080] Distinguishing highly similar objects based on the non-visual attribute of distance can effectively avoid visual similarity misjudgment and effectively improve recognition accuracy and generalization ability. At the same time, distance, as a universal attribute, is applicable not only to similar objects, but also to different types of objects. This makes the method of using distance factors to improve the robustness of classification networks have a wider application prospect.

Claims

1. A classification method for distinguishing target objects with high similarity, characterized in that: Specifically include the following: Step 1: Define a set of high similarity categories; Step 2: Use the YOLO v3 target detection framework to screen and identify objects with high similarity categories; Step 3: Group pre-training Siamese twin neural network; Step 4: Use the Siamese twin network trained in step 3 to perform target classification and output the label classification results; Step 5: Distinguish objects with high similarity based on the label classification results; The Siamese twin neural network described in step 3 consists of two identical sub-networks, each with the same structure and weights, which are used to process paired inputs and learn the similarities and differences between paired images; The idea of ​​meta-learning is used to train the Siamese twin neural network, enabling it to accurately determine whether the pixel ratio of the main part in two images has changed; the steps are as follows: Step 3.1: Dataset preprocessing: Based on the ImageNet dataset, the dataset is divided into training and test sets in a ratio of 8:

2. The two datasets are divided into two groups for training and testing respectively; one group changes the pixel size of the main part of the image and is labeled as 1; The other group is an image with unchanged pixel ratio, and its label is 0. The pixel ratio is used to describe the ratio of the number of object pixels in the image to the total number of pixels in the image, which is the ratio of the object to the total image area. Step 3.2: Construct a Siamese twin neural network: Use a convolutional neural network as the basic structure of the Siamese twin neural network; the mathematical model of the Siamese twin network is as follows: Among them, θ is the model parameter, N is the number of samples in the data set, and L i () represents the loss function of the i-th pair of samples, and Represent the sample sets used for training and testing respectively; Step 3.3: Define the loss function: Based on the binary classification problem, use the binary cross entropy loss function to measure the difference between the Siamese twin neural network prediction results and the true label; the loss function expression is as follows: Where N is the number of samples in the dataset, y i is the true label of the i-th pair of samples, p i is the label predicted by the Siamese twin network; The loss function measures the difference between the Siamese twin network's predictions for the input images and their true labels. When the twin network's predictions for a pair of input images are the same as their true labels, the loss function value is small. Conversely, when the predictions are different from the true labels, the loss function value is large. By minimizing the loss function, the Siamese twin network will gradually learn how to determine whether the pixel ratios of objects in the two images have changed, and achieve more accurate prediction results by adjusting parameters. Step 3.4: Follow the above steps to pre-train the Siamese twin network in groups: use the paired images as the input of the Siamese twin network, calculate the loss function, and finally use the backpropagation algorithm to update the network parameters; Step 5 is as follows: Step 5.1: Obtain the Siamese twin network classification result label: If the classification result label is 1, it means that the object pixel ratio has changed significantly; otherwise, the label is 0, which means that the object pixel ratio has not changed significantly. Step 5.2: Based on the perspective projection effect, the closer an object is to the observer in three-dimensional space, the larger its projection is in the two-dimensional image, and the farther the object is in three-dimensional space, the smaller its projection is in the two-dimensional image. If the pixel ratio of an object in images at different positions does not change, it means that the object is farther away from the device sensor. Conversely, a significant change in the pixel ratio indicates that the object is closer to the device sensor. Successfully distinguish highly similar objects based on the non-visual attribute of distance.

2. A classification method for distinguishing target objects with high similarity according to claim 1, characterized in that: The high similarity categories mentioned in step 1 are specifically: pairs or groups of categories with similar features or attributes in the dataset, which are highly similar in color, shape, and texture; The high-similarity category set includes error-prone categories, identical categories, and similar categories; The error-prone category: objects that are very similar to other objects in appearance, shape, or function and are easy to be confused or mistaken; The same category: multiple objects belong to the same category, and objects in this category have the same characteristics or functions; Similar categories: objects in the same category have appearance features of similar shape, texture or color.

3. The classification method for distinguishing target objects with high similarity according to claim 1, characterized in that: Step 2 is as follows: The device uses sensors onboard to capture environmental images at different locations during its motion. These environmental source images at different device locations are then fed into the YOLO v3 object detection framework. The YOLO v3 object detection framework outputs the location, category, and confidence level of the detected target object. The detection results of each target object are screened, requiring that its confidence level is higher than a preset threshold and that the target object category belongs to a high-similarity category set, thereby filtering out easily confused target information.

4. A classification method for distinguishing target objects with high similarity according to claim 3, characterized in that: The device is a terminal and intelligent system with target detection requirements; the sensor carried by the device is an instrument for capturing images and videos, and is used to implement target detection applications; Acquire environmental images at different positions during the motion process. The distance intervals between the positions of environmental images are dynamically adjusted, and the length of the intervals is determined by the target detection output results of the environmental source image acquired at the previous position. If the environmental source image detection shows that an object belongs to a high-similarity object set, the frequency of environmental acquisition is increased; otherwise, the frequency of environmental acquisition is reduced.

5. The classification method for distinguishing target objects with high similarity according to claim 1, characterized in that: Step 4 is as follows: The trained Siamese twin network is used for the target object classification task of the target detection framework; by screening the images of the same easily confused object information in the environmental source images at different positions in step 2, the same object images at different positions are intercepted and input into the Siamese twin network in pairs, and the classification results are output; if the pixel ratio of the main part in the two images changes greatly, the output label is 1; if the pixel ratio of the main part changes little, the output label is 0.

Images