Infrared image target detection method, device, equipment and medium
By using DarkNet53 backbone network and pyramid bottom-up feature network in infrared image target detection, the problem of complex infrared image target detection algorithm and high computing resource consumption is solved, efficient and accurate target detection is achieved, and suitable for air defense tasks.
Patent Information
- Application Number
- CN202410985592.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-11-15
AI Technical Summary
The existing infrared image object detection algorithm is complex, the computing resources are consumed very much, and it is difficult to deploy on hardware with limited computing resources, especially in environments where target detection efficiency is high but computing power is limited in air defense tasks.
The lightweight and efficient DarkNet53 backbone network is adopted to combine the pyramid bottom-up feature network and decoupling head to train the target detection network through infrared image data sets to reduce the calculation amount of the target detection model and improve detection efficiency.
It realizes efficient target detection on hardware with limited computing resources, ensures detection accuracy, and is easy to deploy, and is suitable for target identification and tracking of air defense weapon systems.
Smart Images

Figure CN118968240A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of target detection technology, and in particular to an infrared image target detection method, device, equipment and medium. Background Art
[0002] In air defense missions, air target identification is crucial. Different detection methods provide air defense systems with a variety of identification and tracking methods. Radar technology has advantages such as a wide search range and fast search speed, but its target positioning accuracy is lower than that of optoelectronic technology, and using radar as an active launch system may expose the position of the air defense system itself.
[0003] Among optoelectronic technologies, visible light technology performs well in image resolution, but is limited by lighting conditions and cannot effectively detect targets at night or in poor visibility. In contrast, infrared imaging has emerged in modern warfare, especially for the detection and tracking of aerial targets such as small drones. Its significant advantages are that targets are more prominent and easier to detect when the sky is the background, and the background is simple and less disturbing. However, the current algorithms for infrared image target detection are complex and require a lot of computing resources, which makes it difficult to deploy on hardware with limited computing resources. Summary of the invention
[0004] The purpose of this application is to provide an infrared image target detection method, device, equipment and medium, which can reduce the calculation amount of the target detection model and improve the target detection efficiency.
[0005] To achieve the above objectives, this application provides the following solutions:
[0006] In a first aspect, the present application provides an infrared image target detection method, comprising:
[0007] An infrared image data set is used to train a target detection network, and the trained target detection network is used as a target detection model; the target detection network includes a backbone network, a neck, and a head connected in sequence; the backbone network uses DarkNet53, the neck uses a pyramid bottom-up feature network, and the head uses a decoupling head;
[0008] The infrared image to be detected is input into the target detection model, and the target detection result is output.
[0009] Optionally, the backbone network includes stem, Dark2, Dark3, Dark4 and Dark5 connected in sequence, the stem is used to perform convolution processing on the input image, and Dark2, Dark3, Dark4 and Dark5 are all composed of convolution blocks and residual blocks.
[0010] Optionally, the neck includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a first upsampling module, a second upsampling module, a first splicing module, a second splicing module, a third splicing module, a fourth splicing module, a first C3 module, a second C3 module, a third C3 module and a fourth C3 module; the first C3 module, the second C3 module, the third C3 module and the fourth C3 module are all used for feature extraction and fusion through three convolution layers and multiple Bottleneck blocks; each Bottleneck block includes two convolution layers connected in series;
[0011] The input of the first convolution module is connected to the output of Dark5, the output of the first convolution module is connected to the input of the first upsampling module, the input of the first splicing module is respectively connected to the output of Dark4 and the output of the first upsampling module, the output of the first splicing module is connected to the input of the first C3 module, the output of the first C3 module is connected to the input of the second convolution module, the output of the second convolution module is respectively connected to the input of the second upsampling module and the input of the third splicing module, the input of the second splicing module is respectively connected to the output of Dark3 and the output of the second upsampling module, the output of the second splicing module is connected to the input of the second C3 module, and the second The output of the C3 module is connected to the input of the third convolution module, the output of the third convolution module is connected to the input of the third splicing module, the output of the third splicing module is connected to the input of the third C3 module, the output of the third C3 module is connected to the input of the fourth convolution module, the output of the fourth convolution module is connected to the input of the fourth splicing module, and the output of the fourth splicing module is connected to the input of the fourth C3 module; the output of the second C3 module serves as the first output of the neck, the output of the third C3 module serves as the second output of the neck, and the output of the fourth C3 module serves as the third output of the neck; the first output, the second output and the third output are all connected to the input of the decoupling head.
[0012] Optionally, the decoupling head includes three prediction branches with the same structure, namely a first prediction branch, a second prediction branch and a third prediction branch, the input of the first prediction branch is connected to the first output, the input of the second prediction branch is connected to the second output, and the input of the third prediction branch is connected to the third output.
[0013] Optionally, the target detection network also includes an output module, which is used to convert the outputs of the first prediction branch, the second prediction branch and the third prediction branch into prediction results respectively, and splice the three prediction results to obtain the target detection result.
[0014] Optionally, when the infrared image data set is used to train the target detection network, each infrared image in the infrared image data set is preprocessed before being input into the target detection network;
[0015] The infrared images are preprocessed, specifically including: resizing the infrared images; performing mean filtering on the resized infrared images; and performing histogram equalization on the mean filtered infrared images to obtain preprocessed infrared images.
[0016] Optionally, in the process of training the target detection network using the infrared image dataset, a cosine annealing algorithm is used to adjust the learning rate, and a stochastic gradient descent algorithm is used to update the parameters of the target detection network.
[0017] In a second aspect, the present application provides an infrared image target detection device, comprising:
[0018] The target detection network training module is used to train the target detection network using the infrared image data set, and use the trained target detection network as the target detection model; the target detection network includes a backbone network, a neck and a head connected in sequence; the backbone network adopts DarkNet53, the neck adopts a pyramid bottom-up feature network, and the head adopts a decoupling head;
[0019] The target detection model application module is used to input the infrared image to be detected into the target detection model and output the target detection result.
[0020] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any one of the above-described infrared image target detection methods.
[0021] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the above-described infrared image target detection methods.
[0022] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0023] The present application provides an infrared image target detection method, device, equipment and medium. The backbone network adopts a lightweight and efficient DarkNet53 network, and the neck adopts a pyramid bottom-up feature network to fuse high-level technical features and low-level technical features, thereby ensuring the accuracy of target detection. This improves the detection efficiency while ensuring the accuracy of target detection. In addition, the lightweight target detection model is easy to deploy on hardware with limited computing resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0025] Figure 1 A flowchart of an infrared image target detection method provided in one embodiment of the present application;
[0026] Figure 2 A schematic diagram of an application scenario of a target detection model provided in an embodiment of the present application;
[0027] Figure 3 A schematic diagram of the structure of a backbone network and a neck in a target detection network provided in an embodiment of the present application;
[0028] Figure 4 A schematic diagram of the structure of the neck and output modules in the target detection network provided in one embodiment of the present application;
[0029] Figure 5 A schematic diagram of the DarkNet53 network structure provided in an embodiment of the present application;
[0030] Figure 6 A schematic diagram of a residual block structure provided in an embodiment of the present application;
[0031] Figure 7 A schematic diagram of a convolution block structure provided in one embodiment of the present application;
[0032] Figure 8 A schematic diagram of the target detection model application process provided in one embodiment of the present application;
[0033] Fig. 9 A schematic diagram of functional modules of an infrared image target detection device provided in another embodiment of the present application.
[0034] Fig.10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0036] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0037] The present application embodiment provides a method for detecting a red image target. Figure 1 As shown, a red image target detection method includes:
[0038] Step 101: Use the infrared image data set to train a target detection network, and use the trained target detection network as a target detection model; the target detection network includes a backbone network, a neck, and a head connected in sequence; the backbone network uses DarkNet53, the neck uses a pyramid bottom-up feature network, and the head uses a decoupled head.
[0039] Step 102: input the infrared image to be detected into the target detection model, and output the target detection result.
[0040] The detection targets of the target detection model in this embodiment include aircraft, and the aircraft include drones, that is, the present application is applicable to target detection of small aircraft. The infrared images in the infrared image data set are infrared images containing various small aircraft (such as drones), and the aircraft are marked in the infrared images. The marked aircraft are label data of the infrared images, and the target detection result is whether the infrared image to be detected contains an aircraft or not.
[0041] In order to improve the accuracy of target detection, when using an infrared image dataset to train a target detection network, each infrared image in the infrared image dataset is preprocessed before being input into the target detection network.
[0042] Preprocessing each infrared image specifically includes: resizing the infrared image; performing mean filtering on the resized infrared image; performing histogram equalization on the mean filtered infrared image to obtain a preprocessed infrared image. Image denoising and image enhancement processing are also included before resizing the infrared image.
[0043] The infrared image to be detected is the infrared image to be detected after preprocessing.
[0044] Adjusting the size of the infrared image is to set the infrared image data to x w,h (0≤w≤width,0≤h≤height), where x w,h is the pixel at the hth row and wth column of the infrared image, width is the image width, and height is the image width.
[0045] The infrared image after resizing is output after mean filtering: Where k is the filter size, N is the number of pixels within the filter window, and mean x,y is the pixel at the yth row and xth column of the infrared image after mean filtering.
[0046] Perform histogram equalization on the infrared image after mean filtering to improve the contrast, let p r (r k ) indicates that the gray level in the input image is r k The probability distribution function of the pixel, its cumulative distribution function is s k , then the expression after histogram equalization is For each input image pixel gray level r k , apply the above formula to get the gray level h of the output image pixel x,y As the input of the target detection network. x,y is the pixel at the yth row and xth column of the infrared image after histogram equalization.
[0047] In an exemplary embodiment, the target detection network is as follows: Figure 3 and Figure 4 As shown, the backbone network includes stem, Dark2, Dark3, Dark4 and Dark5 connected in sequence. The stem is used to perform convolution processing on the input image, specifically converting the spatial information of the input tensor into channel information, and then processing it through a convolution layer. Dark2, Dark3, Dark4 and Dark5 are all composed of convolution blocks and residual blocks.
[0048] In an exemplary embodiment, the DarkNet53 structure is as follows Figure 5 As shown in the figure, after Dark2, Dark3, Dark4 and Dark5, DarkNet53 also includes an average pooling layer (Avgpool), a fully connected layer (Connected) and a Softmax function connected in sequence. Figure 5 Among them, 1000 are the most identifiable object types. Figure 5 The residual block (Residul) is as follows Figure 6 As shown, the convolutional block (Convolutional) is as follows Figure 7 As shown in Figure 1, the convolution block includes a two-dimensional convolution module (Conv2d), a normalization operation (BN), and a LeakyRelu activation function connected in sequence. The formula for the normalization operation is: γ and β are the parameters obtained after training, and y and x are the output and input of the normalization operation.
[0049] The neck includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a first upsampling module, a second upsampling module, a first splicing module, a second splicing module, a third splicing module, a fourth splicing module, a first C3 module, a second C3 module, a third C3 module and a fourth C3 module; the first C3 module, the second C3 module, the third C3 module and the fourth C3 module are all used for feature extraction and fusion through three convolution layers and multiple Bottleneck blocks; each Bottleneck block includes two convolution layers connected in series.
[0050] The first splicing module, the second splicing module, the third splicing module, and the fourth splicing module all splice multiple tensors together along a specified dimension.
[0051] Each C3 module (the first C3 module, the second C3 module, the third C3 module and the fourth C3 module) includes a first convolutional layer, a second convolutional layer, a third convolutional layer, n series-connected Bottleneck blocks and a fifth splicing module. The inputs of the first convolutional layer and the second convolutional layer are connected as the input of the C3 module. The output of the first convolutional layer is connected to the input of the fifth splicing module through n series-connected Bottleneck blocks. The output of the second convolutional layer is connected to the input of the fifth splicing module. The output of the fifth splicing module is connected to the input of the third convolutional layer. The output of the third convolutional layer is the output of the C3 module. The Bottleneck block includes a fourth convolutional layer and a fifth convolutional layer connected in sequence.
[0052] Each C3 module performs feature extraction and fusion through three convolutional layers and multiple Bottleneck blocks.
[0053] The input of the first convolution module is connected to the output of Dark5, the output of the first convolution module is connected to the input of the first upsampling module, the input of the first splicing module is respectively connected to the output of Dark4 and the output of the first upsampling module, the output of the first splicing module is connected to the input of the first C3 module, the output of the first C3 module is connected to the input of the second convolution module, the output of the second convolution module is respectively connected to the input of the second upsampling module and the input of the third splicing module, the input of the second splicing module is respectively connected to the output of Dark3 and the output of the second upsampling module, the output of the second splicing module is connected to the input of the second C3 module, and the second The output of the C3 module is connected to the input of the third convolution module, the output of the third convolution module is connected to the input of the third splicing module, the output of the third splicing module is connected to the input of the third C3 module, the output of the third C3 module is connected to the input of the fourth convolution module, the output of the fourth convolution module is connected to the input of the fourth splicing module, and the output of the fourth splicing module is connected to the input of the fourth C3 module; the output of the second C3 module serves as the first output of the neck, the output of the third C3 module serves as the second output of the neck, and the output of the fourth C3 module serves as the third output of the neck; the first output, the second output and the third output are all connected to the input of the decoupling head.
[0054] The decoupling head includes three prediction branches with the same structure, namely a first prediction branch, a second prediction branch and a third prediction branch, the input of the first prediction branch is connected to the first output, the input of the second prediction branch is connected to the second output, and the input of the third prediction branch is connected to the third output.
[0055] The first prediction branch, the second prediction branch and the third prediction branch all include a fifth convolution module, a sixth convolution module (cls_conv), a seventh convolution module (reg_conv), an eighth convolution module (cls_preds), a ninth convolution module (reg_preds), a prediction module (obj_preds) and a sixth splicing module.
[0056] The output of the fifth convolution module is connected to the input of the sixth convolution module (cls_conv) and the seventh convolution module (reg_conv), respectively. The output of the sixth convolution module is connected to the input of the eighth convolution module. The output of the seventh convolution module is connected to the input of the ninth convolution module and the prediction module, respectively. The outputs of the eighth convolution module, the ninth convolution module and the prediction module are all connected to the input of the sixth splicing module, and the output of the sixth splicing module is used as the output of the prediction branch. The fifth convolution module is used to perform initial convolution processing on the input image to extract features. The sixth convolution module contains two convolution layers for extracting features and performing category classification tasks. The seventh convolution module is used to extract features for target box regression from the feature map. The eighth convolution module is used for predicting the convolution layer of the target category. The ninth convolution module is used for the convolution layer of the target box prediction, and outputs the coordinate information of the target box. The prediction module is used to predict the confidence of the existence of the target and output the prediction result of the confidence of the existence of the target. The sixth splicing module splices multiple tensors together along the specified dimension.
[0057] The target detection network also includes an output module, which is used to convert the outputs of the first prediction branch, the second prediction branch and the third prediction branch into prediction results respectively, and concatenate the three prediction results to obtain the target detection result.
[0058] The output module includes three prediction box conversion modules (get_output_and_grid), namely the first prediction box conversion module, the second prediction box conversion module and the third prediction box conversion module. The input of the first prediction box conversion module is connected to the output of the first prediction branch, the input of the second prediction box conversion module is connected to the output of the second prediction branch, and the input of the third prediction box conversion module is connected to the output of the third prediction branch. The output module also includes the seventh splicing module and the output layer (output_det). The prediction box conversion module is used to convert the network output into the coordinates of the prediction box and return the grid of the prediction box. The output layer is used to output the detection result.
[0059] The outputs of the first prediction box conversion module, the second prediction box conversion module and the third prediction box conversion module are all connected to the input of the seventh splicing module, and the output of the seventh splicing module is connected to the input of the output layer.
[0060] During the training process, the target detection network also includes a prediction boundary box category and loss determination module (get_assignments), a loss value determination module (cal_loss), a label information storage module (labels) and a loss value output module (output_loss). The prediction boundary box category and loss determination module and the input of the loss value determination module are connected to the output of the seventh splicing module, the label information storage module is connected to the prediction boundary box category and loss determination module input, the prediction boundary box category and loss determination module output is connected to the loss value determination module input, and the loss value determination module output is connected to the loss value output module.
[0061] The predicted bounding box category and loss determination module is used to perform the matching process in target detection, matching the predicted bounding box with the true bounding box to determine the target category and loss of each predicted bounding box. The loss value determination module is used to calculate the loss value of the target detection model. The label information storage module is used to store the true label information of each sample, and the loss value output module is used to output the loss value.
[0062] In the process of training the target detection network using the infrared image dataset, the cosine annealing algorithm is used to adjust the learning rate.
[0063] Where i is the index value of the warm restart after the learning rate has been executed for several epochs (dividing all samples by the size of batchsize) (where batchsize represents the number of samples trained at one time), and Represent the maximum and minimum values of the learning rate, respectively, and define the range of the learning rate, which remains unchanged after each restart. cur Indicates how many epochs are currently executed. i Indicates the epoch number when the algorithm is run for the i-th time. n t Represents the current learning rate.
[0064] The stochastic gradient descent algorithm is used to update the parameters of the target detection network, where the update formula is: Among them, θ represents the target detection network parameters, t represents the current number of iterations, η represents the learning rate, and controls the size of the step. i is the loss function for the i-th data point or data batch.
[0065] The red image target detection method provided in this application can be applied to Figure 2 In the application environment shown.
[0066] The deployment of the target detection model specifically includes: saving the target detection model parameters and writing them into the SD card on the computer side. When the target detection model is applied, the target detection model parameters in the SD card are read by the microcontroller (MCU) to load the target detection model parameters in the SD card. Specifically, the target detection model parameters in the SD card are loaded into the double rate synchronous dynamic random access memory (DDR) of the hardware platform in batches. After confirming that the parameter loading is complete, the computer side transmits an infrared image data (an image gray value array) through the network port connected to the field programmable gate array (FPGA) side and saves it in the DDR of the hardware platform. After all inputs are loaded, the MCU will receive feedback and release computing tasks to the FPGA according to the network structure. The FPGA calculates each layer of the network according to the computing instructions. First, the FPGA reads the weight parameters required for this layer in the DDR and the input feature map of this layer (infrared image gray value matrix or the network output of the previous layer). The above process is iterated until all network layers are calculated. Through element concatenation and transposition operations, the original input image will eventually be cut into a 80×80×(Class+Bbox+Objectness) feature map and saved in the DDR. Class represents the number of categories. Since only people are recognized, it is 1. Bbox is the predicted rectangular box parameter. The rectangular box parameter includes t x ,t y ,t w and t h , t x and t y represents the offset relative to the feature map grid, t w and t h Indicates the size of the rectangular box, and Objectness indicates the probability of whether the grid has a target. Finally, all grids need to be traversed to find the grid with the highest target score. The score calculation formula is score = Objectness × Class. FPGA transmits the Bbox parameters of the grid with the highest score to the computer through the network port for analysis, and displays it on the computer monitor after marking in the original image to realize target recognition of the pre-processed infrared image. The application process is as follows Figure 8 shown.
[0067] Figure 2 The hardware platform includes memories including DDR and dynamic random access memory (DRAM), and DDR3 is a model of DDR.
[0068] The convolution calculation engine obtains some feature maps and convolution kernel parameter data from the on-chip buffer and completes some convolution operations ( Figure 7) to ensure complete accuracy. After the calculation is completed, the data is passed to the quantization unit. The quantization unit quantizes the result according to the network model quantization parameter (scale, offset quntization = (input-offset) * scale) to ensure consistent calculation accuracy. The pooling calculation unit is an optional unit to complete the downsampling of the feature map. Its data flow can flow directly into the on-chip buffer or enter the on-chip buffer after pooling calculation.
[0069] Based on the application of target detection model, the detected target is tracked and located, the target position and motion trajectory are determined, and effective target recognition is provided for the air defense weapon system. This application discloses a red image target detection method that can better adapt to the target recognition and tracking requirements under different environmental conditions, thereby improving the combat effectiveness of the air defense weapon system and having high practical value.
[0070] Based on the same inventive concept, the embodiment of the present application also provides a red image target detection device for implementing the red image target detection method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more red image target detection device embodiments provided below can refer to the limitations of the red image target detection method above, and will not be repeated here.
[0071] In an exemplary embodiment, Fig. 9 As shown, a red image target detection device is provided, comprising:
[0072] The target detection network training module 201 is used to train the target detection network using the infrared image data set, and use the trained target detection network as the target detection model; the target detection network includes a backbone network, a neck and a head connected in sequence; the backbone network adopts DarkNet53, the neck adopts a pyramid bottom-up feature network, and the head adopts a decoupling head.
[0073] The target detection model application module 202 is used to input the infrared image to be detected into the target detection model and output the target detection result.
[0074] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store red image target detection data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a red image target detection method is implemented.
[0075] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.
[0076] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0077] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0078] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0079] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0080] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a data processing logic of a programmable logic device, etc., but is not limited thereto.
[0081] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0082] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A method for infrared image target detection, It is characterized in that The infrared image target detection method comprises: An infrared image data set is used to train a target detection network, and the trained target detection network is used as a target detection model; the target detection network includes a backbone network, a neck, and a head connected in sequence; the backbone network uses DarkNet53, the neck uses a pyramid bottom-up feature network, and the head uses a decoupling head; The infrared image to be detected is input into the target detection model, and the target detection result is output.
2. The infrared image target detection method according to claim 1, It is characterized in that The backbone network includes stem, Dark2, Dark3, Dark4 and Dark5 connected in sequence, the stem is used to perform convolution processing on the input image, and Dark2, Dark3, Dark4 and Dark5 are all composed of convolution blocks and residual blocks.
3. The infrared image target detection method according to claim 2, It is characterized in that The neck includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a first upsampling module, a second upsampling module, a first splicing module, a second splicing module, a third splicing module, a fourth splicing module, a first C3 module, a second C3 module, a third C3 module and a fourth C3 module; the first C3 module, the second C3 module, the third C3 module and the fourth C3 module are all used for feature extraction and fusion through three convolution layers and multiple Bottleneck blocks; each Bottleneck block includes two convolution layers connected in series; The input of the first convolution module is connected to the output of Dark5, the output of the first convolution module is connected to the input of the first upsampling module, the input of the first splicing module is respectively connected to the output of Dark4 and the output of the first upsampling module, the output of the first splicing module is connected to the input of the first C3 module, the output of the first C3 module is connected to the input of the second convolution module, the output of the second convolution module is respectively connected to the input of the second upsampling module and the input of the third splicing module, the input of the second splicing module is respectively connected to the output of Dark3 and the output of the second upsampling module, the output of the second splicing module is connected to the input of the second C3 module, and the second The output of the C3 module is connected to the input of the third convolution module, the output of the third convolution module is connected to the input of the third splicing module, the output of the third splicing module is connected to the input of the third C3 module, the output of the third C3 module is connected to the input of the fourth convolution module, the output of the fourth convolution module is connected to the input of the fourth splicing module, and the output of the fourth splicing module is connected to the input of the fourth C3 module; the output of the second C3 module serves as the first output of the neck, the output of the third C3 module serves as the second output of the neck, and the output of the fourth C3 module serves as the third output of the neck; the first output, the second output and the third output are all connected to the input of the decoupling head.
4. The infrared image target detection method according to claim 3, It is characterized in that The decoupling head includes three prediction branches with the same structure, namely a first prediction branch, a second prediction branch and a third prediction branch, the input of the first prediction branch is connected to the first output, the input of the second prediction branch is connected to the second output, and the input of the third prediction branch is connected to the third output.
5. The infrared image target detection method according to claim 4, It is characterized in that The target detection network also includes an output module, which is used to convert the outputs of the first prediction branch, the second prediction branch and the third prediction branch into prediction results respectively, and splice the three prediction results to obtain the target detection result.
6. The infrared image target detection method according to claim 1, It is characterized in that When using the infrared image data set to train the target detection network, each infrared image in the infrared image data set is preprocessed before being input into the target detection network; The infrared images are preprocessed, specifically including: resizing the infrared images; performing mean filtering on the resized infrared images; and performing histogram equalization on the mean filtered infrared images to obtain preprocessed infrared images.
7. The infrared image target detection method according to claim 1, It is characterized in that In the process of training the target detection network with infrared image dataset, the cosine annealing algorithm is used to adjust the learning rate, and the stochastic gradient descent algorithm is used to update the parameters of the target detection network.
8. An infrared image target detection device, It is characterized in that The infrared image target detection device comprises: The target detection network training module is used to train the target detection network using the infrared image data set, and use the trained target detection network as the target detection model; the target detection network includes a backbone network, a neck and a head connected in sequence; the backbone network adopts DarkNet53, the neck adopts a pyramid bottom-up feature network, and the head adopts a decoupling head; The target detection model application module is used to input the infrared image to be detected into the target detection model and output the target detection result.
9. A computer device, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the infrared image target detection method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the infrared image target detection method described in any one of claims 1 to 7 is implemented.
Citation Information
Patent Citations
Apple leaf disease detection method based on improved YOLOX-S algorithm
CN116363517A
Complex road scene target detection method based on YOLOX model
CN117237612A
Small target detection model training method, small target detection method and related equipment
CN117746015A
Cited By
Infrared image splicing and labeling method and system based on image fusion and medium
CN119152163A
