Infrared image conversion method and device suitable for visible light target recognition model

Abstract

The invention discloses an infrared image conversion method suitable for a visible light target recognition model, and the method comprises the steps: training an image processing network model which is used for converting an infrared image into a visible light image in the following mode: inputting infrared image sample data into the image processing network model, and obtaining a converted image sample, and inputting the converted image sample into a feature extraction network in a trained visible light target recognition model, carrying out target feature extraction, and inputting the extracted target feature sample into a trained first target classifier and a trained image classifier respectively, and taking the image classification loss function value of the image classifier and the target classification loss function value of the first target classifier as supervision loss function values, and adjusting model parameters of the image processing network model until the supervision loss function values reach the expectation. The method is suitable for the infrared image under the condition that the visible light target recognition model is kept unchanged.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image-based target recognition, in particular, to an infrared image conversion method suitable for a visible light target recognition model. BACKGROUND

[0002] A visible light target recognition model for visible light image target recognition is trained by using easily-obtained visible light image sample data, and if the model is used to recognize targets in infrared images, the recognition performance of the visible light target recognition model will obviously decrease because infrared images have no color information and are quite different from visible light images.

[0003] Because the model parameters of the visible light target recognition model cannot be easily changed, in order to make the visible light target recognition model also applicable to infrared images, the existing processing methods mostly adopt an image conversion method, that is, a mapping between infrared images and visible light images is established by using the data of visible light images and infrared images, and then the infrared images are converted into visible light images based on the established mapping, but such conversion has limited effect on improving the performance of the visible light target recognition model, and the converted images are different from the real visible light images, and the recognition effect of the visible light target recognition model on the converted images is still unsatisfactory. SUMMARY

[0004] The present application provides an infrared image conversion method suitable for a visible light target recognition model, so as to improve the target recognition effect of infrared images while keeping the model parameters of the visible light target recognition model unchanged.

[0005] The present application provides an infrared image conversion method suitable for a visible light target recognition model, so as to improve the target recognition effect of infrared images while keeping the model parameters of the visible light target recognition model unchanged.

[0006] The infrared image sample data is input into the image processing network model to obtain converted image samples,

[0007] The converted image samples are input into a feature extraction network in the trained visible light target recognition model to perform target feature extraction,

[0008] The extracted target feature samples are respectively input into a first target classifier trained for target classification based on visible light images and an image classifier trained for classifying visible light images and infrared images,

[0009] The image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier are used as the supervised loss function value, and the model parameters of the image processing network model are adjusted,

[0010] The iteration is performed until the supervised loss function value reaches the expectation.

[0011] In the embodiment of the present application, the trained visible light target recognition model is trained by using visible light image sample data;

[0012] The image classifier is trained in the following manner:

[0013] The feature extraction network in the visible light target recognition model is used to perform target feature extraction on the visible light image sample data and the infrared image sample data respectively, and the extracted target feature samples are input into the image classifier to train the image classifier.

[0014] In the embodiment of the present application, the visible light image sample data and the infrared image sample data are image sample data after geometric size normalization of target images;

[0015] The first target classifier is trained in the following manner:

[0016] The extracted visible light target feature samples are input into the first target classifier to train the first target classifier, and a trained first target classifier is obtained,

[0017] Or,

[0018] The feature extraction network in the visible light target recognition model is used to perform target feature extraction on the visible light image sample data, and the extracted target feature samples are input into the first target classifier to train the first target classifier, and a trained first target classifier is obtained.

[0019] In the embodiment of the present application, the image classification loss function value is determined in the following manner:

[0020] The value 1 is subtracted from the confidence value of the visible light classification result in the classification result output by the image classifier to obtain the image classification loss function value.

[0021] In the embodiment of the present application, the supervised loss function value is determined in the following manner: the image classification loss function value and the target classification loss function value are weighted and summed to obtain a fusion loss function value.

[0022] In the embodiment of the present application, the image processing network model is a machine learning model,

[0023] The parameters of the feature extraction network remain unchanged.

[0024] The second aspect of the application provides a target recognition method, which comprises:

[0025] In the case that the image to be recognized is a visible light image, the image to be recognized is input into the trained visible light target recognition model to perform visible light image target recognition,

[0026] In the case that the image to be recognized is an infrared image, the image to be recognized is converted by using the trained image processing network model for converting infrared images into visible light images, and the converted image is input into the trained visible light target recognition model to perform infrared image target recognition,

[0027] The image processing network model is obtained according to the steps of any one of the infrared image conversion methods suitable for the visible light target recognition model.

[0028] The image processing network model is obtained according to the steps of any one of the infrared image conversion methods suitable for the visible light target recognition model.

[0029] In the embodiment of the application, the trained visible light target recognition model is obtained by training using visible light image sample data.

[0030] The third aspect of the application provides an infrared image conversion device suitable for a visible light target recognition model, which comprises:

[0031] An image processing network model for converting infrared images into visible light images,

[0032] The image processing network model in the image processing network model is trained in the following manner:

[0033] Infrared image sample data is input into the image processing network model to obtain converted image samples,

[0034] The converted image samples are input into the feature extraction network of the trained visible light target recognition model to extract target features, and the extracted target feature samples are respectively input into a first target classifier trained for target classification based on visible light images and an image classifier trained for classifying visible light images and infrared images.

[0035] The image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier are used as the supervision loss function value, and the model parameters of the image processing network model are adjusted,

[0036] The above steps are repeatedly performed until the supervision loss function value reaches the expected value.

[0037] The fourth aspect of the present application provides a target recognition device, which comprises:

[0038] a target recognition module, configured to perform target recognition on the to-be-recognized image when the to-be-recognized image is a visible light image, and perform target recognition on the image converted by the image processing network model when the to-be-recognized image is an infrared image,

[0039] an image processing network model, configured to convert the infrared image into a visible light image to obtain a converted image, and send the converted image to the target recognition module;

[0040] wherein,

[0041] the image processing network model is trained in the following manner:

[0042] the infrared image sample data is input into the image processing network model to obtain a converted image sample,

[0043] the converted image sample is input into a feature extraction network in the trained visible light target recognition model to perform target feature extraction, and the extracted target feature sample is input into a first target classifier trained for target classification based on a visible light image and an image classifier trained for image classification,

[0044] the image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier are used as the supervision loss function value, and the model parameters of the image processing network model are adjusted,

[0045] the above steps are repeatedly performed until the supervision loss function value reaches an expectation.

[0046] The infrared image conversion method suitable for a visible light target recognition model improved by the embodiments of the present application uses a trained image classifier and a trained target classifier to train an image processing network model for converting infrared images into visible light images, so that the trained image processing network model converts infrared images into visible light images suitable for the visible light target recognition model, and under the condition that the model parameters of the visible light target recognition model remain unchanged, the infrared image and the visible light image share the same visible light target recognition model, thereby expanding the capability of the visible light target recognition model for target recognition of infrared images. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 FIG. 1 is a flowchart of an infrared image conversion method suitable for a visible light target recognition model according to an embodiment of the present application.

[0048] Figure 2 Another flowchart of the infrared image conversion method suitable for the visible light target recognition model.

[0049] Figure 3 An example of the image classifier training is shown.

[0050] Figure 4 An example of the target classifier training is shown.

[0051] Figure 5 An example of the image processing network model for the infrared image conversion training is shown.

[0052] Figure 6 An example of the target recognition of the present application is shown.

[0053] Figure 7 An example of the target recognition device is shown.

[0054] Figure 8 Another example of the target recognition device or the infrared image conversion device suitable for the visible light target recognition model is shown. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical means and advantages of the present application more clear, the present application is further described in detail below with reference to the accompanying drawings.

[0056] The applicant finds that, in order to improve the performance of the existing visible light target recognition model with fixed model parameters on the infrared image, the existing method mainly focuses on the visual similarity of the infrared image converted into the visible light image, without combining the visible light target recognition model itself, so that the target recognition accuracy of the infrared image is limited to improve. The present application embodiment proposes an infrared image conversion method suitable for a visible light target recognition model, which not only makes the infrared image close to the visual effect of the visible light image, but also greatly improves the target recognition accuracy of the infrared image by the visible light target recognition model.

[0057] Referring to Figure 1 shown, Figure 1 An example of the infrared image conversion method suitable for the visible light target recognition model of the present application is shown. The image processing network model for converting the infrared image into the visible light image is trained in the following manner:

[0058] Step 101, inputting the infrared image sample data into the image processing network model to be trained to obtain the converted image sample,

[0059] As an example, the image processing network model can be any kind of machine learning model, such as UNet model, etc.

[0060] Step 102, input the converted image sample into the feature extraction network in the trained visible light target recognition model to perform target feature extraction,

[0061] As an example, the trained visible light target recognition model is a model obtained by training with visible light image sample data, which at least includes a feature extraction network for target feature extraction and a target classifier for target recognition. To avoid confusion with the target classifier used for training the image processing network model below, the target classifier in the trained visible light target recognition model is referred to as the second target classifier.

[0062] In this step, the converted image sample is subjected to target feature extraction by the feature extraction network in the trained visible light target recognition model, wherein the parameters of the feature extraction network remain unchanged.

[0063] Step 103, input the extracted target feature samples into the trained first target classifier for target classification based on visible light images and the trained image classifier for classifying visible light images and infrared images, respectively, to obtain sample target classification results output by the first target classifier and sample image classification results output by the image classifier,

[0064] As an example, the first target classifier can be determined according to the training requirements of the image processing network model and can be trained in the following manner:

[0065] The target feature extraction network in the trained visible light target recognition model is used to perform target feature extraction on the visible light image sample data, and the extracted target feature samples are input into the first target classifier to train the first target classifier.

[0066] As an example, the trained image classifier can be obtained in the following manner:

[0067] The target feature extraction network in the trained visible light target recognition model is used to perform target feature extraction on the visible light image sample data and the infrared image sample data, and the extracted target feature samples are input into the image classifier to train the image classifier.

[0068] It should be understood that the training process of the first target classifier and the training process of the image classifier each includes a process of calculating a loss function value according to the difference between the output result of the classifier and the expected result, and adjusting the model parameters of the classifier according to the loss function value.

[0069] In the training process of the image classifier and the first target classifier, the parameters of the feature extraction network remain unchanged.

[0070] Step 104, taking the image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier as the supervised loss function value, adjusting the model parameters of the image processing network model, repeatedly executing steps 101-104 until the supervised loss function value reaches the expectation.

[0071] As an example, the image classification loss function value is determined in the following manner:

[0072] Subtracting the confidence value of the visible light classification result in the sample classification result output by the image classifier from the value 1 obtains the image classification loss function value.

[0073] The supervised loss function value is determined in the following manner: summing the image classification loss function value and the target classification loss function value by weighting to obtain a fusion loss function value.

[0074] The infrared image conversion method suitable for the visible light target recognition model provided by the embodiments of the present application combines the trained image classifier and the trained first target classifier with the image processing network model to be trained, and supervises and trains the image processing network model by fusing the loss function values of the image classifier and the first target classifier, so that the infrared image can be converted into a visible light image suitable for the existing visible light target recognition model. By using the image processing network model, the target recognition performance of the visible light image is ensured to be unchanged under the condition that the model parameters of the visible light target recognition model remain unchanged, and the target recognition accuracy of the infrared image by the visible light target recognition model reaches a similar result as the visible light, which extends the visible light target recognition model to the infrared image.

[0075] For the convenience of understanding the embodiments of the present application, the following will be described in combination with a face recognition model, and it should be understood that the present application is not limited to the face recognition model.

[0076] Referring to Figure 2 , it is shown that Figure 2 Another flowchart of the infrared image conversion method suitable for the visible light target recognition model.

[0077] Step 201, inputting the infrared image sample data and the visible light image sample data into a feature extraction network in any existing visible light target recognition model to extract target feature samples, and inputting the extracted target feature samples into an image classifier for classifying infrared images and visible light images to train the image classifier.

[0078] The existing visible light target recognition model is trained by visible light image sample data, that is, the existing visible light target recognition model is a trained visible light target recognition model, which includes a feature extraction network and a second target classifier connected with the feature extraction network for target recognition.

[0079] As an example, visible light and infrared image sample data containing the same target are collected, for example, for face recognition, the data of 100 people can be collected, and 50 visible light images and 50 infrared images of each person are collected.

[0080] The target in the image sample is geometrically normalized, for example, for the collected face image sample, the position of the face and the position of the facial feature points in the face can be obtained by face detection, for example, the face detector RetinaFace (RetinaFace: Single-stage Dense Face Localisation in the Wild) can be used; then a geometric normalization method is used to normalize the geometric size of the face image sample, for example, according to the positions of the left eye center, right eye center, nose tip, left corner of the mouth, and right corner of the mouth in the face, and the reference positions of the corresponding five points, the face image sample is normalized to a set size, wherein the visible light image sample and the infrared image sample are subjected to the same geometric normalization processing.

[0081] The normalized image sample is input into the feature extraction network in the visible light target recognition model to extract target features, for example, the feature extraction network in the trained face recognition model is used to extract face features, and then an image classifier is trained based on the extracted face feature samples.

[0082] The image classifier can use a commonly used softmax classifier in deep learning, which is a 2-class classifier, the visible light image is one class and the infrared image is another class.

[0083] Referring to Figure 3 It is shown that Figure 3 An example of training of the image classifier is shown. The infrared image sample data and the visible light image sample data are randomly input into the feature extraction network, and the target features extracted by the feature extraction network are input into the image classifier.

[0084] Step 202, input the visible light image sample into the feature extraction network in the visible light target recognition model to extract target feature samples, wherein the model parameters of the feature extraction network remain unchanged, the extracted visible light target feature samples are input into the first target classifier, or the visible light target feature samples extracted in step 201 are directly input into the first target classifier to train the first target classifier, and a retrained target classifier is obtained.

[0085] For example, the visible light image sample after geometric size normalization is input into the feature extraction network in the face recognition model to extract face features; then a face ID (Identity) classifier is trained based on the extracted face features. The face ID classifier can use a softmax classifier commonly used in deep learning, and all images of a person correspond to a separate class, that is, each class of the classifier corresponds to the ID of a person, and the number of classes classified by the classifier can be determined according to the training requirements of the image processing network model. There are many improved methods for the loss function of the first target classifier, such as ArcFace loss (Additive Angular Margin Loss for Deep Face Recognition) as the face ID classifier of face identity classification.

[0086] Referring to Figure 4 , a schematic diagram of training of the first target classifier is shown. Figure 4 The visible light image sample data is input into the feature extraction network, and the target features extracted by the feature extraction network are input into the first target classifier to train the first target classifier.

[0087] Step 203, on the basis of the trained image classifier and the first target classifier, an image processing network model for converting infrared images into visible light images is trained, so that the converted images can be used for target recognition by the target recognition model.

[0088] Referring to Figure 5 , a schematic diagram for training the image processing network model for converting infrared images is shown. Figure 5 The infrared image sample is input into the image processing network model to be trained, wherein the infrared image sample is a geometric size normalized image sample; the image sample output by the image processing network model is input into the feature extraction network in the visible light target recognition model to extract target features, and the model parameters of the feature extraction network remain unchanged; the target features extracted by the feature extraction network are input into the image classifier and the first target classifier respectively, wherein the model parameters of the classifiers remain unchanged, and a plurality of loss values are used as supervision signals in the image processing network training process, specifically including: target classification loss value, and image classification loss value, wherein:

[0089] The network structure that the second target classifier needs to learn in the training of the general target recognition model is different from that of the first target classifier. Here, the model parameters of the first target classifier are fixed, coming from the first target classifier trained in step 202. During training, the infrared image sample of the same target is processed by the image processing network model and input to the feature extraction network and the first target classifier as the visible light image sample of the target, and the first target classifier should have the same output result in theory. At this time, the target classification loss function value should be relatively small, and thus the target classification loss function value can be used as a supervision loss function value for the entire network training.

[0090] After the infrared image sample is processed by the image processing network model, if the converted image sample is very suitable for the subsequent visible light target recognition model, the converted image sample is classified by using the image classifier trained in step 201 after the target feature is extracted. At this time, the sample classification result will change. The original classification result is the infrared image category, and now it should be the visible light image category, that is, by adjusting the model parameters of the image processing network model, the infrared image can be converted into a visible light image as suitable as possible for the visible light target recognition model. Therefore, the confidence of the visible light image classification result in the sample classification result in the image classifier (denoted as score_RGB) is processed, and 1-score_RGB is used as the loss function value of the image classifier as a supervision loss function value for the entire network training.

[0091] In order to obtain the optimal result, the target classification loss function value and the image classification loss function value can be further fused to supervise the training of the entire network. The fused loss function value L can be represented as:

[0092] L = Loss TAR x W TAR + Loss RGB x W RGB,

[0093] wherein W TAR and W RGB are weighting coefficients, which can be set by experience to be a more reasonable weighting coefficient to better optimize the network, Loss TAR represents the target classification loss function value, and Loss RGB represents the image classification loss function value.

[0094] For example, the infrared face image normalized in terms of geometric size is input to the image processing network model, and the image processing network model is used to complete the conversion of the infrared image, and the feature extraction network (fixed parameters) in the existing face recognition model is used to complete the extraction of the face feature.

[0095] In order to make the converted image suitable for the existing face recognition model, during training, the infrared face image of the same person is processed by the image processing network and is sent into the face recognition model as the visible light image of the person, and the face ID classifier should have the same output, and the loss function value Loss_id of the ID classification at this time should be relatively small.

[0096] After the infrared face image of a person is processed by the image processing network model, if the converted image is very suitable for the subsequent visible light face recognition model, then after the face recognition features are extracted from the converted image, the image classifier is used for classification, and the sample classification result will change at this time. The original classification result is the infrared category, and now it should be the visible light category.

[0097] The fused loss L can be expressed as:

[0098] L = Loss_id * W_id + Loss_RGB * W_RGB

[0099] Wherein, W_id and W_RGB are weighting coefficients, and Loss_id is the loss function value of the face ID classifier.

[0100] During the training process, the model parameters of the image processing network model are adjusted according to the supervised loss function value, and the process is repeated until the supervised loss function value reaches the expectation.

[0101] Referring to Figure 6 As shown in the figure, Figure 6 is a schematic diagram of target recognition of the embodiment of the present application. After the image processing network model is trained, if the to-be-identified image is a visible light image, the to-be-identified image is directly input into the visible light target recognition model used when the image processing network model is trained, so as to perform visible light image target recognition, and if the to-be-identified image is an infrared image, the to-be-identified image is converted by using the trained image processing network model, and the converted image is input into the visible light target recognition model, so as to perform infrared image target recognition. In this way, not only the performance of the target recognition model for the visible light image is unchanged, but also the target recognition accuracy of the target recognition model for the infrared image reaches a result similar to that of the visible light image.

[0102] For example, during deployment of the face recognition model, the visible light image and the infrared image are subjected to the same face detection and geometric size normalization processing; if the image is a visible light image, the image is directly sent into the face recognition model; and if the image is an infrared image, the image is first sent into the trained image processing network model to perform conversion processing, and then the converted image is sent into the face recognition model.

[0103] The embodiment of the present application also provides a schematic diagram of an infrared image conversion device suitable for a visible light target recognition model, and the device comprises:

[0104] an image processing network model for converting an infrared image into a visible light image,

[0105] wherein the image processing network model is trained in the following manner:

[0106] inputting infrared image sample data into the image processing network model to obtain a converted image sample,

[0107] inputting the converted image sample into a feature extraction network in the trained visible light target recognition model to perform target feature extraction, and inputting the extracted target feature sample into a first target classifier trained for target classification based on a visible light image and an image classifier trained for image classification of visible light images and infrared images,

[0108] adjusting model parameters of the image processing network model based on the image classification loss function value of the image classifier and the target classification loss function value of the first target classifier as a supervision loss function value,

[0109] repeating the above steps until the supervision loss function value reaches an expectation, and deploying the trained image processing network model in, for example, a server.

[0110] Referring to Figure 7 shown, Figure 7 is a schematic diagram of a target recognition device, which comprises:

[0111] a target recognition module for performing target recognition on a to-be-recognized image when the to-be-recognized image is a visible light image, and performing target recognition on a converted image obtained by the image processing network model module when the to-be-recognized image is an infrared image,

[0112] an image processing network model module for converting an infrared image into a visible light image to obtain a converted image, and sending the converted image to the target recognition module.

[0113] Referring to Figure 8 shown, Figure 8 is another schematic diagram of a target recognition device or an infrared image conversion device applicable to a visible light target recognition model, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the steps of the target recognition method or the steps of the infrared image conversion method applicable to the visible light target recognition model.

[0114] The memory can include a random access memory (RAM) and can also include a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0115] The processor described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc. It can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0116] The embodiment of the present application further provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of target identification of the embodiment of the present application or the steps of the infrared image conversion method suitable for the visible light target identification model.

[0117] For the device / network side equipment / storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts refer to the part of the method embodiment.

[0118] In this document, the terms "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising", or any other variations thereof are intended to cover non-exclusive inclusions, so that a process, method, article, or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article, or device. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of additional identical elements in the process, method, article, or device including the element.

[0119] The above description is only the preferred embodiment of the present application and is not intended to limit 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 infrared image conversion method suitable for visible light target recognition models, characterized in that, The image processing network model used to convert infrared images into visible light images is trained as follows: Infrared image sample data is input into the image processing network model to obtain converted image samples. The converted image samples are input into the feature extraction network of the trained visible light target recognition model to extract target features. The extracted target feature samples are input into a first target classifier trained for target classification based on visible light images, and a trained image classifier, respectively. The image classifier is used to classify visible light images and infrared images. The image processing network model parameters are adjusted using the image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier. Repeat the process until the value of the supervised loss function reaches the expected value.

2. The infrared image conversion method as described in claim 1, characterized in that, The trained visible light target recognition model is obtained by training with visible light image sample data; The image classifier is trained in the following manner: Using the feature extraction network in the visible light target recognition model, target features are extracted from visible light image sample data and infrared image sample data respectively. The extracted target feature samples are then input into the image classifier to train the image classifier.

3. The infrared image conversion method as described in claim 2, characterized in that, The visible light image sample data and infrared image sample data are image sample data after geometric size normalization of the target image; The first target classifier is trained in the following manner: The extracted visible light target feature samples are input into the first target classifier to train the first target classifier, resulting in the trained first target classifier. or, Using the feature extraction network in the visible light target recognition model, target features are extracted from visible light image sample data. The extracted target feature samples are then input into the first target classifier to train the first target classifier, resulting in a trained first target classifier.

4. The infrared image conversion method as described in claim 1, characterized in that, The image classification loss function value is determined in the following manner: Subtracting the confidence value of the visible light classification result from the image classifier's output from the value 1 yields the image classification loss function value.

5. The infrared image conversion method as described in claim 1, characterized in that, The supervised loss function value is determined as follows: the image classification loss function value and the target classification loss function value are weighted and summed to obtain the fusion loss function value.

6. The infrared image conversion method according to any one of claims 1 to 5, characterized in that, The image processing network model is a machine learning model. The parameters of the feature extraction network remain unchanged.

7. A target recognition method, characterized in that, The method includes: When the image to be identified is a visible light image, the image is input into a trained visible light target recognition model to perform visible light image target recognition. When the image to be identified is an infrared image, a trained image processing network model for converting infrared images into visible light images is used to convert the image. The converted image is then input into a trained visible light target recognition model for infrared image target recognition. in, The image processing network model is obtained according to the steps of an infrared image conversion method applicable to a visible light target recognition model as described in any one of claims 1 to 6.

8. The target recognition method as described in claim 7, characterized in that, The trained visible light target recognition model is obtained by training with visible light image sample data.

9. An infrared image conversion device suitable for visible light target recognition models, characterized in that, include: An image processing network model is used to convert infrared images into visible light images. The image processing network model within the image processing network model is trained in the following manner: Infrared image sample data is input into the image processing network model to obtain converted image samples. The converted image samples are input into the feature extraction network of a trained visible light target recognition model to extract target features. The extracted target feature samples are then input into a first target classifier trained for target classification based on visible light images, and a trained image classifier, which is used to classify visible light and infrared images. The image processing network model parameters are adjusted using the image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier. Repeat the process until the value of the supervised loss function reaches the expected value.

10. A target recognition device, characterized in that, The device includes: The target recognition module is used to perform target recognition on the image to be recognized when the image to be recognized is a visible light image, and to perform target recognition on the image after it has been transformed by the image processing network model when the image to be recognized is an infrared image. An image processing network model is used to convert infrared images into visible light images, obtain the converted images, and send the converted images to the target recognition module; in, The image processing network model is trained in the following manner: Infrared image sample data is input into the image processing network model to obtain converted image samples. The converted image samples are input into the feature extraction network of a trained visible light target recognition model to extract target features. The extracted target feature samples are then input into a first target classifier trained for target classification based on visible light images, and a trained image classifier, which is used to classify visible light and infrared images. The image processing network model parameters are adjusted using the image classification loss function value of the sample image classification result output by the image classifier and the target classification loss function value of the sample target classification result output by the first target classifier. Repeat the process until the value of the supervised loss function reaches the expected value.

Images