An army armor fine-grained identification method and system based on multi-granularity fusion

Abstract

The application discloses an army armor fine-grained recognition method and system based on multi-granularity fusion, which can start from three angles of original images, target objects and discriminative parts of the target objects, comprehensively detect results of models of all angles, and distinguish sub-categories by comprehensively considering overall, local and subtle differences, so that more accurate classification results are obtained. First, an army armor target image dataset is acquired and preprocessed. A multi-granularity fusion classification model is constructed, including a coarse-grained classification model, a medium-grained classification model and a fine-grained classification model. The coarse-grained classification model is used for classifying original images in the image dataset. The medium-grained classification model is used for extracting target objects in the images for classification. The fine-grained classification model is used for extracting discriminative parts in the images for classification. The prediction scores of the three are weighted, and finally the detection score of the image is obtained, and the sub-category of the image is further recognized. The trained multi-granularity fusion recognition model is used for classifying images.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and relates to a method and system for fine-grained recognition of army armor based on multi-granularity fusion. BACKGROUND

[0002] With the development of artificial intelligence technology, information technology is widely used in military battlefields. Battlefield target recognition has become an important means of battlefield reconnaissance. Image target recognition is an important problem in the field of computer vision. In recent years, for the coarse-grained target recognition task, such as identifying whether the reconnaissance target is a tank or a military truck, a relatively good recognition effect can be achieved. However, for the fine-grained target recognition task, such as distinguishing whether the reconnaissance target is an M1A3 main battle tank or an RDF / LT light tank, so as to formulate the most effective action plan according to the characteristics of the armor type, it is still difficult for machines. However, in actual combat operations, the land battlefield armor target is the focus of unmanned reconnaissance platforms and strikes, and realizing fine-grained recognition can more effectively serve combat operations.

[0003] Therefore, there is currently a lack of a classification scheme for military battlefield army armor targets that can integrate the overall, local and subtle differences, and it is impossible to achieve accurate classification. SUMMARY

[0004] Therefore, the application provides a method and system for fine-grained recognition of army armor based on multi-granularity fusion, which can start from three angles of original images, target objects and discriminative parts of target objects, integrate the detection results of models of various angles, integrate overall, local and subtle differences to distinguish subcategories, and thus obtain more accurate classification results.

[0005] To achieve the above purpose, the technical scheme of the application is as follows: a method for fine-grained recognition of army armor based on multi-granularity fusion, comprising the following steps:

[0006] Step 1: Obtain an army armor target image data set.

[0007] Step 2: Perform data preprocessing on the image data in the army armor target image data set.

[0008] Step 3: Construct a multi-granularity fusion classification model, including three classification models under different granularities, namely a coarse-grained classification model, a medium-grained classification model and a fine-grained classification model.

[0009] The coarse-grained classification model is used for classifying the original images in the image data set and retaining the original features in the images, to obtain a coarse-grained classification result.

[0010] The medium-grained classification model is used for extracting target objects in the images and classifying the target objects, to obtain a medium-grained classification result.

[0011] The fine-grained classification model is used to extract discriminative parts in the image, and the fine-grained classification result is obtained according to the discriminative parts.

[0012] The multi-granularity fusion classification model is trained by using the army armored target image training data set to obtain a trained multi-granularity fusion recognition model.

[0013] Step four: obtaining the army armored target image test data, taking it as input, and inputting it into the trained multi-granularity fusion recognition model.

[0014] Step five: the coarse-grained, medium-grained and fine-grained classification results are weighted and fused to obtain the final detection result of the image.

[0015] Further, step one is specifically: first, obtaining an image data set composed of multiple images of army ground military armored targets shot by a reconnaissance vehicle and its label, using stratified sampling method to divide the data set, and randomly extracting 80% of the images in each category as training image data set and the remaining 20% of the images as test image data set.

[0016] Further, step two, data preprocessing includes the following steps: adjusting the size of the image data in the army armored target image data set. The image data after size adjustment is subjected to data enhancement, and the data enhancement methods include up-down flipping, left-right rotation, enhancing / reducing picture brightness and mirror data enhancement method. The image data after data enhancement is divided into multiple candidate image sub-blocks, and the candidate image sub-blocks with a correlation greater than a threshold with the category label are selected to form a candidate sub-block image set.

[0017] Further, the coarse-grained classification model is specifically: taking the candidate sub-block image set as training data, and its training label is the corresponding sub-category; loading the network structure and weight parameters of the ImageNet pre-trained Resnet101 model, using the above training data set to fine-tune the model, thereby obtaining the coarse-grained classification model.

[0018] Further, the medium-grained classification model is specifically: taking the candidate sub-block image set obtained by data preprocessing as training data, and its training label is the corresponding sub-category. The candidate sub-block image set is preprocessed, first, the class activation mapping CAM method is adopted to obtain the saliency map of the target object in the image, the saliency map is binarized and the connected region is extracted to obtain the target region of the object in the picture, so as to obtain the target object in the candidate sub-block image, which is the target object training data set. Taking the target object training data set and its corresponding sub-class label as input data, inputting it into the trained Resnet101 model for fine-tuning to generate a classification model about the target object, which is called medium-grained classification model.

[0019] Further, the fine-grained classification model, specifically: first, the candidate sub-block image set obtained by data preprocessing is taken as input data, second, the part clustering algorithm is used to generate the cluster taking the part as the category, and the class label obtained by manual labeling is replaced; finally, the clustered candidate sub-block image set and its class label are input into the pre-trained Resnet101 model for fine-tuning to generate the classification model about the object part, which is called fine-grained classification model.

[0020] Further, step five, the coarse-grained, medium-grained and fine-grained classification results are weighted and fused to obtain the final detection result of the image, and the specific steps are as follows:

[0021] The test image is input into the trained coarse-grained, medium-grained and fine-grained classification model, and the classification results of the three different-grained classification models are calculated respectively as:

[0022] Score coarse-grain , Score middle-grain , Score fine-grain ;

[0023] The final result Score final is obtained by weighting, and the calculation method is as follows:

[0024] Score final =α×Score coarse-grain +β×Score middle-grain +γ×Score fine-grain

[0025] Wherein, alpha, beta and gamma are the weights of the prediction scores of the three different-grained models in the total score, and the range is set to be between 0-1, and the step is 0.1.

[0026] Further, after step five, the final detection result of the image is also evaluated, and the evaluation index of the detection result, i.e. the accuracy Acc, is constructed:

[0027]

[0028] Wherein R c represents the number of correct detection in all samples, and R represents the number of all samples.

[0029] Another embodiment of the present application also provides an army armor fine-grained recognition system based on multi-grained fusion, which comprises an army armor target image data acquisition module, a data preprocessing module, a multi-grained fusion classification model module and a test result evaluation module.

[0030] An army armored target image data acquisition module is configured to acquire an army armored target image dataset, which is classified into a training image dataset and a test image dataset.

[0031] A data preprocessing module is configured to perform data preprocessing on image data in the army armored target image dataset.

[0032] A multi-granularity fusion classification model module is configured to construct a multi-granularity fusion classification model, which includes three classification models of different granularities, namely a coarse-grained classification model, a medium-grained classification model, and a fine-grained classification model. The coarse-grained classification model is configured to classify original images in the image dataset and retain original features in the images to obtain coarse-grained classification results. The medium-grained classification model is configured to extract target objects in the images and classify the target objects to obtain medium-grained classification results. The fine-grained classification model is configured to extract discriminative parts in the images and classify the discriminative parts to obtain fine-grained classification results. The multi-granularity fusion classification model is trained using the army armored target image training dataset to obtain a trained multi-granularity fusion recognition model. The army armored target image test data is acquired and input into the trained multi-granularity fusion recognition model. The coarse-grained, medium-grained, and fine-grained classification results are weighted and fused to obtain the final detection result of the image.

[0033] A test result evaluation module is configured to evaluate the final detection result of the image and construct an evaluation index for the detection result, namely precision Acc: wherein R c represents the number of correctly detected samples, and R represents the total number of samples.

[0034] Beneficial effects:

[0035] 1. The army armored fine-grained recognition method based on the attention mechanism is adopted. First, ground military armored target image data collected by an unmanned reconnaissance vehicle and an unmanned reconnaissance aircraft is acquired. Second, a classification model based on multi-granularity fusion is constructed. The model mainly fuses three classification models of different granularities. The coarse-grained classification model focuses on retaining original features of the images, the medium-grained classification model focuses on locating target objects in the images, and the fine-grained classification model focuses on discriminative parts of the target objects. The three models extract overall, local, and subtle features from different angles to train the fine-grained classification model for the army armored image, thereby obtaining a trained multi-granularity fusion classification model. Third, the test dataset is input into the trained multi-granularity fusion classification model to obtain detection scores of the three classification models. The prediction scores of the three models are weighted to obtain the final detection score of the image, which further identifies the subcategory of the image to assist the operator in making decisions for armored target recognition in the land battle field.

[0036] 2、The application increases the medium granularity classification module, can effectively eliminate the influence of image background noise, and further promotes the model to learn more effective target object features. The model mainly obtains the saliency map of the target object through the class activation map CAM (class activation map) method, then binarizes it, extracts the connected region, thereby obtaining the target object in the candidate sub-block image, and then takes the target object image and the sub-class label as input data to train the medium granularity classification model.

[0037] 3、The existing method basically adopts the method of labeling the parts of the sub-class, which leads to a huge workload of data labeling and wastes a lot of labor cost. In order to avoid this problem, the application adopts the part clustering method to realize automatic labeling. First, the part sub-block image set in the second part data preprocessing is obtained; secondly, the part clustering algorithm is adopted to generate the cluster with the part as the category for the discriminative part data set, instead of manual labeling; finally, the clustered sub-block image set and the class label are input into the pre-trained Resnet101 model for fine-tuning to obtain better classification performance. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the overall flowchart of the application. DETAILED DESCRIPTION

[0039] The application will be described in detail below with reference to the drawings and examples.

[0040] The application provides a land army armor fine-grained recognition method based on multi-granularity fusion, and the flowchart is as shown in Figure 1 The specific steps are as follows:

[0041] Step one: obtaining a land army armor target image data set.

[0042] This step is mainly responsible for obtaining the data set. For the application, first, a ground military armor image data set and its label photographed by a plurality of reconnaissance vehicles are obtained. In order to ensure that the training set and the test set maintain the same distribution as possible, a stratified sampling method is adopted to divide the data set, and 80% of the images in each category are randomly extracted as the training data set, and the remaining 20% of the images are taken as the test data set.

[0043] Step two: data preprocessing.

[0044] This step is mainly to preprocess the obtained land army armor target image data set. The main steps of data preprocessing are:

[0045] S201: adjusting the picture size, adjusting the picture size to a size suitable for the model, such as 224*224;

[0046] S202: Data augmentation, by flipping up and down, rotating left and right, enhancing / reducing picture brightness, mirroring, etc. Data augmentation methods, existing image data is processed to expand the data set.

[0047] S203: In order to locate the specific position of the target object, the image is usually divided into multiple sub-blocks during data preprocessing, and the method adopted by the present application is selective search algorithm, so as to generate candidate sub-block images and expand the data set. The detailed algorithm process is as follows:

[0048]

[0049] S204: The candidate image sub-blocks are generated by the above algorithm, but the image data set obtained contains some noise. Therefore, the present application first uses a pre-trained Resnet101 model to fine-tune on the training set of the present application, thereby obtaining an effective classification filtering model. Then, the picture is input into the trained classification filtering model, and a threshold is set. The candidate images greater than the threshold are retained, which are considered to have strong relevance to a certain category; and the candidate images less than the threshold are filtered, which are considered to have weak relevance to a certain category. Thus, the candidate sub-block image set with strong relevance is further obtained from the candidate image set.

[0050] Step three: build a multi-granularity fusion classification model.

[0051] This step mainly builds three classification models with different granularities. The coarse-grained classification model pays more attention to the original image, the medium-grained classification model mainly focuses on positioning the target object, and the fine-grained classification model mainly focuses on positioning the discriminative part of the target object. From these three angles, the results of each angle model detection are weighted and fused to obtain the final detection result of the image. This model integrates the overall, local and subtle differences to distinguish sub-categories. The present application details the specific implementation of the three models as follows:

[0052] Coarse-grained classification model

[0053] This module mainly focuses on classifying the original image, aiming to retain the original features in the image. The candidate sub-block data set with strong relevance is used to fine-tune the pre-trained Resnet101 model, thereby obtaining better coarse-grained image classification performance. The specific steps are as follows:

[0054] Sa301: The candidate sub-block data set with strong relevance obtained in the second part of data preprocessing is used as training data, and the training label is the corresponding sub-category.

[0055] Sa302: load the network structure and weight parameters of the ImageNet pre-trained Resnet101 model, use the above training data set, fine-tune the model to obtain a coarse-grained classification model.

[0056] Medium-grained classification model

[0057] The module mainly focuses on positioning the target object in the image, and the existing weak supervision method is mostly dedicated to mining the discriminative parts of the target object, while ignoring the positioning of the target object, resulting in excessive noise. In order to avoid this influence, the present application adds a medium-grained classification module, which can effectively eliminate the influence of image background noise and further promote the model to learn more effective target object features. The present model mainly uses the class activation mapping CAM method to first obtain the saliency map of the target object, then binarizes it, extracts the connected region, and thus obtains the target object in the candidate sub-block image, and then takes the target object image and its sub-class label as input data to train the medium-grained classification model. The specific steps are as follows:

[0058] Sb301: select the selective search algorithm and filter the generated candidate sub-block data set with strong relevance as training data in the second part of data preprocessing, and the training label is the corresponding sub-class.

[0059] Sb302: preprocess the candidate sub-block data set, first use the class activation mapping CAM method to obtain the saliency map of the target object in the image.

[0060] Sb303: binarize and extract the connected region of the saliency map to obtain the target region of the object in the picture, so as to obtain the target object in the candidate sub-block image, that is, the target object training data set.

[0061] Sb304: take the target object training data set and its corresponding sub-class label as input data, and input it into the trained Resnet101 model for fine-tuning, which can generate a classification model about the target object, called a medium-grained classification model.

[0062] Fine-grained classification model

[0063] The module mainly focuses on positioning the discriminative parts in the target object, and aims to distinguish sub-categories from local and subtle differences. Existing methods basically adopt the method of labeling the parts of sub-categories, resulting in a huge workload of data labeling and wasting a lot of labor costs. In order to avoid this problem, the present application adopts the part clustering method to realize automatic labeling here. First, the part sub-block image set in the second part data preprocessing is obtained; secondly, the part clustering algorithm is adopted to generate the cluster of the discriminative part data set with the part as the category, instead of manual labeling; finally, the clustered sub-block image set and its class label are input into the pre-trained Resnet101 model for fine-tuning to obtain better classification performance. The specific steps are as follows:

[0064] Sc301: the candidate sub-block data set with strong relevance obtained in the second part data preprocessing is taken as input data;

[0065] Sc302: since the candidate part sub-block set in the above step is disordered and not semantically aligned, there may be multiple head and tail parts in different perspectives. In view of the importance of different parts in the classification result, the present application adopts the clustering method to cluster the discriminative candidate part sub-block set, cluster the sub-blocks of the same part into a category, and generate part class labels. Wherein, the discriminative part sub-block set of picture i is Xi={xi1, xi2, …, xim}. The K-means clustering method is adopted to cluster the part sub-block set Xi into k categories A1, A2, …, Ak (the value of k can be optimized by cross-validation), align the parts, and generate class labels. The K-means clustering process is as follows:

[0066]

[0067] Sc303: the k cluster part sub-block data set after clustering and its corresponding class label are taken as training data and input into the trained Resnet101 model for fine-tuning to generate a classification model about object parts, called fine-grained classification model.

[0068] Step four: obtain the test data of the army armored target image.

[0069] This step is mainly responsible for obtaining test data, mainly for data preprocessing of the test data set divided in the first part, adjusting the picture size, taking it as input, and inputting it into the trained multi-granularity fusion recognition model.

[0070] Step five: fuse the results of the multi-granularity classification model.

[0071] This step mainly fuses the results of the three different granularity classification models to obtain the final score corresponding to the image. The specific steps are as follows:

[0072] The test image is respectively input into the trained coarse-grained, medium-grained and fine-grained classification models, and the prediction score results Score coarse-grain , Score middle-grain , Score fine-grain

[0073] The final result is obtained by weighting, and the calculation method is as follows:

[0074] Score final =α*Score coarse-grain +β*Score middle-grain +γ*Score finne-grain

[0075] Wherein, alpha, beta, gamma are the weights of the prediction scores of the three different granularity models in the total score, and the range is set to be between 0-1, and the step is 0.1. This value can be obtained by the grid search method to obtain the optimal combination value.

[0076] The three different granularity classification models focus on different levels of complementary features, and the overall, local and subtle features are fused, so that the prediction performance of the whole fused multi-granularity classification model is improved.

[0077] Step six: test result evaluation index.

[0078] This part is mainly aimed at the evaluation index of the test result, and the evaluation index in the present application is mainly the accuracy (Accuracy), and the calculation formula is as follows:

[0079]

[0080] Wherein, R c represents the number of correct detection in all samples, and R represents the number of all samples.

[0081] The present application also provides an army armor fine-grained recognition system based on multi-granularity fusion, which comprises an army armor target image data acquisition module, a data preprocessing module, a multi-granularity fusion classification model module and a test result evaluation module.

[0082] The army armor target image data acquisition module is used for acquiring army armor target image data set, which is classified into training image data set and test image data set.

[0083] The data preprocessing module is used for data preprocessing of image data in the army armor target image data set.

[0084] The multi-granularity fusion classification model module is used for constructing a multi-granularity fusion classification model, which includes three classification models in different granularities, i.e., a coarse-granularity classification model, a medium-granularity classification model and a fine-granularity classification model; the coarse-granularity classification model is used for classifying original images in an image data set and retaining original features in the images to obtain a coarse-granularity classification result; the medium-granularity classification model is used for extracting target objects in the images and classifying the target objects to obtain a medium-granularity classification result; and the fine-granularity classification model is used for extracting discriminative parts in the images and classifying the discriminative parts to obtain a fine-granularity classification result; the multi-granularity fusion classification model is trained by using an army armored target image training data set to obtain a trained multi-granularity fusion recognition model; an army armored target image test data is acquired, which is taken as an input and transmitted into the trained multi-granularity fusion recognition model; and the coarse-granularity, medium-granularity and fine-granularity classification results are weighted and fused to obtain a final detection result of the image.

[0085] The test result evaluation module is used for evaluating the final detection result of the image, constructing an evaluation index of the detection result, i.e., an accuracy Acc: wherein R c represents the number of correct detections in all samples, and R represents the number of all samples.

[0086] To sum up, the above is only a preferred embodiment of the present application, and is not used for limiting the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An armored fine-grained recognition method for the army based on multi-granularity fusion, characterized in that, It comprises the following steps: Step one: obtain the image data set of the army armored target; Step two: data preprocessing is performed on the image data in the army armored target image data set; The data preprocessing comprises the following steps: The size of the image data in the army armored target image data set is adjusted; The data enhancement method includes up-down flipping, left-right rotation, enhancing / reducing picture brightness, and mirror data enhancement method; The data enhanced image data is divided into a plurality of candidate image sub-blocks, and the candidate image sub-blocks with a correlation greater than a threshold with the category label are selected to form a candidate sub-block image set; Step three: a multi-granularity fusion classification model is constructed, including three different granularity classification models, namely a coarse-grained classification model, a medium-grained classification model, and a fine-grained classification model; The coarse-grained classification model is used to classify the original image in the image data set and retain the original features in the image to obtain a coarse-grained classification result; The medium-grained classification model is used to extract the target object in the image and classify according to the target object to obtain a medium-grained classification result; The fine-grained classification model is used to extract the discriminative part in the image and classify according to the discriminative part to obtain a fine-grained classification result; A trained multi-granularity fusion recognition model is obtained by training the multi-granularity fusion classification model using the army armored target image training data set; Step four: obtain the army armored target image test data, input it into the trained multi-granularity fusion recognition model as input; Step five: the coarse-grained, medium-grained and fine-grained classification results are weighted and fused to obtain the final detection result of the image; The test image is input into the trained coarse-grained, medium-grained and fine-grained classification models, and the classification results of the three different granularity classification models are calculated respectively as: Score coarse-grain , Score middle-grain , Score fine-grain ; The final result Score is obtained by weighting final The calculation is as follows: Score final = α × Score coarse-grain + β × Score middle-grain + γ × Score fine-grain Wherein, α, β, γ are the weights of the prediction scores of the three different granularity models in the total score, and the range is set to 0-1, and the step is 0.

1.

2. The method of claim 1, wherein, The step one is specifically: First, obtain the image data set of the army ground military armored target shot by the reconnaissance vehicle and its label, divide the data set by stratified sampling method, and randomly extract 80% of the images in each category as training image data set, and the remaining 20% of the images as test image data set.

3. The method of claim 1, wherein, The coarse-grained classification model is specifically: The candidate sub-block image set is used as training data, and the training label is the corresponding sub-category; the network structure and weight parameters of the ImageNet pre-trained Resnet101 model are loaded, the above training data set is used, and the model is fine-tuned to obtain the coarse-grained classification model.

4. The method of claim 1, wherein, The medium-grained classification model is specifically: The candidate sub-block image set obtained by data preprocessing is used as training data, and the training label is the corresponding sub-category; The candidate sub-block image set is pre-processed. Firstly, a CAM method is adopted to obtain a saliency map of the target object in the image. The saliency map is binarized and connected region extraction is performed to obtain a target region of the object in the image, so as to obtain the target object in the candidate sub-block image, that is, the target object training data set; The target object training data set and the corresponding sub-class label are input into the trained Resnet101 model for fine-tuning to generate a classification model of the target object, referred to as a medium-granularity classification model.

5. The method of claim 1, wherein, The fine-granularity classification model is specifically: Firstly, the candidate sub-block image set obtained by data preprocessing is used as input data, and then a part clustering algorithm is used to generate a cluster of discriminative part data sets with parts as categories, instead of manually labeled class labels; Finally, the clustered candidate sub-block image set and the class label are input into the pre-trained Resnet101 model for fine-tuning to generate a classification model of the object part, referred to as a fine-granularity classification model.

6. The method of claim 1, wherein, After step five, the final detection result of the image is evaluated, and an evaluation index of the detection result, that is, the accuracy Acc, is constructed. where R c represents the number of samples detected correctly, and R represents the number of all samples.

7. An army armor fine-grained recognition system based on multi-granularity fusion, characterized in that, It comprises an army armored target image data acquisition module, a data preprocessing module, a multi-granularity fusion classification model module and a test result evaluation module. The army armored target image data acquisition module is used to obtain an army armored target image data set, which is classified into a training image data set and a test image data set. The data preprocessing module is used to perform data preprocessing on the image data in the army armored target image data set. The data preprocessing comprises the following steps: The size of the image data in the army armored target image data set is adjusted; The size-adjusted image data is subjected to data enhancement, and the data enhancement methods include up-down flipping, left-right rotation, enhancing / reducing picture brightness and mirror data enhancement methods; The data-enhanced image data is divided into a plurality of candidate image sub-blocks, and candidate image sub-blocks with a correlation greater than a threshold with the class label are selected to form a candidate sub-block image set; The multi-granularity fusion classification model module is used to construct a multi-granularity fusion classification model, which comprises three classification models under different granularities, namely a coarse-granularity classification model, a medium-granularity classification model and a fine-granularity classification model; the coarse-granularity classification model is used to classify the original image in the image data set and retain the original features in the image to obtain a coarse-granularity classification result; the medium-granularity classification model is used to extract a target object in the image and classify the target object to obtain a medium-granularity classification result; The fine-granularity classification model is used to extract a discriminative part in the image and classify the discriminative part to obtain a fine-granularity classification result; The multi-granularity fusion classification model is trained by using the army armored target image training data set to obtain a trained multi-granularity fusion recognition model; an army armored target image test data is obtained, which is input into the trained multi-granularity fusion recognition model; The coarse-grained, medium-grained and fine-grained classification results are weighted and fused to obtain the final detection result of the image; The test image is input into the trained coarse-grained, medium-grained and fine-grained classification models respectively, and the classification results of the three different-grained classification models are calculated respectively as follows: Score coarse-grain , Score middle-grain , Score fine-grain ; The final result Score is obtained by weighting final The calculation is as follows: Socre final = α × Score conrSe-grain + β × Score middle-grain + γ × Score fine-grain Wherein, alpha, beta and gamma are the weights of the prediction scores of the three different-grained models in the total score, and the range is set to 0-1 with a step of 0.

1. The test result evaluation module evaluates the final detection result of the image, and constructs an evaluation index of the detection result, i.e., an accuracy Acc: Wherein r c represents the number of correct detections in all samples, and R represents the number of all samples.

Images