High-precision infrared target detection method based on YOLOv9 improvement

Abstract

The invention discloses a high-precision infrared target detection method based on YOLOv9 improvement, and the method comprises the steps: taking a disclosed hot data set as target detection image data, converting the target detection image data into a YOLO training format, and obtaining a training set; the method comprises the following steps: based on a YOLOv9 basic framework, introducing an SDI module, a DySample dynamic sampler and an MPDIOU loss function, and constructing an infrared target detection model; inputting the training set into the infrared target detection model, and performing optimization training on the model to obtain a trained infrared target detection model; and inputting an image to be detected into the trained infrared target detection model, and outputting an infrared target detection result. According to the invention, the target detection precision of the infrared image can be improved. The method can be widely applied to the technical field of image processing.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a high-precision infrared target detection method based on improved YOLOv9. Background Art

[0002] Infrared target detection is an important research direction in the field of computer vision and image processing, and is widely used in military reconnaissance, security monitoring, night vision imaging and other fields.

[0003] In recent years, with the development of deep learning technology, object detection algorithms based on convolutional neural networks (CNN) have made great progress, such as the R-CNN series, YOLO (You Only Look Once) series, and SSD, etc. Among them, the YOLO series of algorithms has received widespread attention for its high efficiency and real-time performance.

[0004] The latest YOLOv9 algorithm has further improved the detection speed while maintaining high accuracy. However, the existing YOLOv9 algorithm still has some shortcomings in infrared image target detection tasks, such as poor detection of small targets, low robustness to complex background environments, and high latency. The algorithm needs to be improved and optimized based on the characteristics of infrared images. Summary of the invention

[0005] In view of this, in order to solve the technical problem that the existing target detection method is not suitable for infrared images, which leads to low detection accuracy in infrared image target detection tasks, the present invention proposes a high-precision infrared target detection method based on YOLOv9 improvement, the method comprising the following steps:

[0006] Use the public thermal dataset as target detection image data and convert it into the YOLO training format to obtain the training set;

[0007] Based on the YOLOv9 basic framework, the SDI module, DySample dynamic sampler and MPDIoU loss function are introduced to build an infrared target detection model;

[0008] Inputting the training set into the infrared target detection model, optimizing the model and obtaining a trained infrared target detection model;

[0009] The image to be tested is input into the trained infrared target detection model, and the infrared target detection result is output.

[0010] In some embodiments, it also includes:

[0011] Data enhancement is performed on the image data in the training set from two aspects: light intensity and rotation direction.

[0012] In some embodiments, the infrared target detection model includes a backbone network and a detection head. The backbone network consists of multiple convolutional layers, pooling layers, and ELAN attention modules. It gradually reduces the input resolution while extracting more and more advanced semantic features, and enhances the feature representation of the target through the attention module to prepare for subsequent detection tasks; the detection head includes an upsampling merging module, a downsampling merging module, an ELAN block, and an output convolutional layer. Through upsampling and feature fusion, semantic features at different levels are integrated to finally obtain the feature representation required for classification or detection.

[0013] The present invention also proposes a high-precision infrared target detection system based on YOLOv9 improvement, the system comprising:

[0014] A data acquisition unit, used to use a public thermal data set as target detection image data and convert it into a YOLO training format to obtain a training set;

[0015] The model building unit, based on the YOLOv9 basic framework, introduces the SDI module, DySample dynamic sampler and MPDIoU loss function to build an infrared target detection model;

[0016] A model training unit, used for inputting the training set into the infrared target detection model, optimizing the model and obtaining a trained infrared target detection model;

[0017] The detection unit is used to input the image to be detected into the trained infrared target detection model and output the infrared target detection result.

[0018] Based on the above scheme, the present invention provides a high-precision infrared target detection method based on the improvement of YOLOv9, and introduces the SDI (Semantic and Detail Infusion) module in the infrared target detection model. Compared with the cascade operation of the serialized feature fusion, the SDI module can better enhance the semantic information and detail information in the image by integrating the hierarchical feature map generated by the encoder; at the same time, the DySample dynamic upsampler is also introduced instead of the traditional kernel-based sampler, which not only improves the model processing efficiency, but also avoids time-consuming dynamic convolution and additional sub-network generation of dynamic kernels, significantly reducing the computational workload and delay. On this basis, the MPDIoU loss function is introduced to more accurately reflect the difference between the predicted frame and the real frame, thereby obtaining a more accurate loss value and improving the computational efficiency. On the whole, the present invention has great improvements in the number of model parameters and detection accuracy, and achieves a balance between lightweight and high performance compared to the original model, ensuring the accuracy of infrared target object detection while also laying a solid foundation for its transplantation to terminal devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the steps of a high-precision infrared target detection method improved based on YOLOv9 of the present invention;

[0020] Figure 2 It is a data flow diagram of the SDI module of the present invention;

[0021] Figure 3 It is a data flow diagram of the DySample dynamic sampler of the present invention;

[0022] Figure 4 It is a structural block diagram of a high-precision infrared target detection system improved based on YOLOv9 in the present invention. DETAILED DESCRIPTION

[0023] 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.

[0024] It should be noted that, for the convenience of description, only the parts related to the relevant invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0025] It should be understood that the "system", "device", "unit" and / or "module" used in this application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0026] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0027] In the description of the embodiments of the present application, "plurality" means two or more than two. The following terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0028] In addition, flow charts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0029] Reference Figure 1 , is a flow chart of an optional example of a high-precision infrared target detection method based on the improved YOLOv9 proposed in the present invention. The method can be applied to a computer device. The detection method proposed in this embodiment may include but is not limited to the following steps:

[0030] Step S1, obtaining image data and constructing a training set;

[0031] Step S2: Based on the YOLOv9 basic framework, the SDI module and DySample dynamic sampler are introduced to build an infrared target detection model;

[0032] Step S3, training the infrared target detection model based on the training set to obtain a trained infrared target detection model;

[0033] Step S4: input the image to be tested into the trained infrared target detection model, and output the infrared target detection result.

[0034] In some feasible embodiments, the step S1 specifically includes:

[0035] Use the corresponding database to load the target detection image data, convert it to the YOLO training format, perform overall image enhancement (light intensity, rotation direction), and finally divide it into training and test sets.

[0036] In this embodiment, the target detection image data is selected from the public FLIRv2 thermal dataset, and the self-made Python code is used to convert it into the YOLO training format and divide it into a training set and a test set. The training set is 8862 images, and the test set is 1366 images. A total of 10228 images.

[0037] Then the whole image is enhanced, specifically from the two aspects of light intensity and rotation direction. In terms of light intensity, the brightness and contrast of the image are randomly adjusted to obtain a partially enhanced image; in terms of rotation direction, the angle of the image is randomly adjusted to obtain the remaining enhanced image.

[0038] In this embodiment, the final enhanced complete training set is 44,310 images, and the test set is 6,830 images, for a total of 51,140 images.

[0039] In some feasible embodiments, the infrared target detection model includes a backbone network and a detection head, wherein:

[0040] The backbone network includes a convolutional layer, a pooling layer and an ELAN module;

[0041] Specifically, the backbone network may include a first downsampling module, a second downsampling module, a first ELAN (Efficient Layer Attention Network) module, a third downsampling network, a second ELAN module, a fourth downsampling module, a third ELAN module, a fifth downsampling module, and a fourth ELAN module;

[0042] The backbone network consists of multiple convolutional layers, pooling layers, and ELAN modules. It gradually reduces the input resolution while extracting more and more advanced semantic features, and enhances the feature representation of the target through the attention module to prepare for subsequent detection tasks.

[0043] The detection head includes an upsampling merging module, a downsampling merging module, an ELAN block and an output convolutional layer;

[0044] Specifically, the detection head may include a first spatial pyramid pooling ELAN module, a first upsampling merging module, a fifth ELAN module, a second upsampling merging module, a sixth ELAN module, a first average pooling convolution downsampling merging module, a seventh ELAN module, a second average pooling convolution downsampling merging module, an eighth ELAN module, a second spatial pyramid pooling ELAN module, a third upsampling merging module, a ninth ELAN module, a fourth upsampling merging module, a tenth ELAN module and a detection module.

[0045] The detection head includes an upsampling module, a feature fusion module, an ELAN module, and a final output convolution layer. Through upsampling and feature fusion, semantic features at different levels are integrated to finally obtain the feature representation required for classification or detection.

[0046] Between the two downsampling merging modules, an SDI module is embedded to replace the Concat operation;

[0047] In the upsampling merging module, a DySample dynamic sampler is introduced.

[0048] In some feasible embodiments, the working process in the backbone network is:

[0049] First, the input image is processed using two consecutive ordinary convolution blocks to perform convolution operations on the input image to extract low-level features while reducing the resolution.

[0050] The extracted features are then passed through the first ELAN module to enhance the representation of the target in the image feature map while suppressing the background.

[0051] Next, the features pass through an average pooling layer and a convolutional layer. The feature map is first average pooled to reduce the resolution, and then the convolutional layer is used to further extract features.

[0052] After that, the features are further enhanced through the ELAN attention mechanism block composed of RepNCSPELAN4. Here, RepNCSPELAN4 is composed of multiple ELAN4 modules, and ELAN4 adjusts the spatial attention module in ELAN1 to a regional enhancement attention module.

[0053] Next, the features pass through an average pooling layer and a convolutional layer to further reduce the resolution while extracting features.

[0054] Finally, the features are again passed through the ELAN attention mechanism block composed of RepNCSPELAN4 to achieve the effect of enhancing the target feature representation for the last time.

[0055] In some feasible embodiments, in the detection head, the working process is as follows:

[0056] First, the SPPELAN module is used to perform spatial pyramid pooling and ELAN attention enhancement on the last layer of features of the backbone network to obtain enhanced semantic features.

[0057] The enhanced semantic features are then upsampled and fused with the next layer of features in the backbone network to obtain a fused feature map.

[0058] The fused feature map then passes through the ELAN module composed of RepNCSPELAN4, and multiple ELAN4 attention modules are used to further enhance the target features.

[0059] The enhanced feature map is upsampled and fused again, and fused with the shallower feature map in the backbone network to obtain a higher resolution fused feature map.

[0060] The fused high-resolution feature map is again passed through the ELAN module composed of RepNCSPELAN4 for attention enhancement.

[0061] Afterwards, the attention-enhanced feature map is downsampled through average pooling and convolutional layers, and feature fused with the previous attention-enhanced feature map to obtain a fused intermediate feature map.

[0062] The fused mid-level feature map is again passed through the ELAN module composed of RepNCSPELAN4 for attention enhancement.

[0063] Then, the attention-enhanced intermediate feature map is downsampled again through average pooling and convolutional layers, and feature fused with the original semantic feature map to obtain a fused high-level feature map.

[0064] The fused high-level feature map is again passed through the ELAN module consisting of RepNCSPELAN4 for attention enhancement.

[0065] Next, the SPPELAN module is repeatedly used to perform spatial pyramid pooling and attention enhancement on the last layer of semantic features output by the backbone network.

[0066] The enhanced semantic features are upsampled again and concatenated with the high-resolution, intermediate, and high-level feature maps obtained by the previous fusion to obtain the enhanced feature map.

[0067] The enhanced feature map is subjected to attention enhancement by the ELAN module composed of RepNCSPELAN4 to obtain an enhanced multi-scale feature representation.

[0068] Similarly, the enhanced features are upsampled again to obtain the final enhanced feature map.

[0069] The final feature maps are respectively subjected to attention enhancement by the ELAN modules composed of RepNCSPELAN4 to obtain enhanced multi-scale feature representation.

[0070] The enhanced multi-scale features are fed into the detection module to detect objects of different scales.

[0071] In some feasible embodiments, the workflow of the SDI module is as follows: Figure 2 As shown:

[0072] 1. Upsampling: Through the upsampling process, the SDI module increases the size of the feature map (I=2) from the lower level to match the size of the current level. This helps bring more detailed information into the current feature map;

[0073] 2. Identity Map: This usually means that the feature map is passed directly to the next operation without any modification. Here, it may mean that the feature map of the third level is passed directly to the SDI module for processing without any changes;

[0074] 3. Downsampling: In contrast to upsampling, this step reduces the size of the feature map of the higher level (I=4) to match the size of the third level. This also brings the semantic information of the higher level into the current level. \

[0075] 4. Application of Attention (SpatialAttention): Apply spatial and channel attention mechanisms to process the features of each level generated by the encoder. This process enables the features to integrate local spatial information and global channel information.

[0076] Using the hierarchical feature maps generated by the encoder, we first apply spatial and channel attention mechanisms to the features at each level i This process enables the features to integrate local spatial information and global channel information as follows:

[0077]

[0078] in, represents the feature map after processing of the i-th layer, and denote the parameters of spatial attention and channel attention of the i-th layer respectively. In addition, we apply 1×1 convolution to The number of channels is reduced to c, where c is a hyperparameter. The resulting feature map is represented as Among them, H i , W i and c represent Width, height and channels.

[0079] Next, we need to send the refined feature maps to the decoder. At each decoder level i, we use as the target reference. We then resize each j-th level feature map to match For the same resolution, the formula is:

[0080] Where D, I and U represent adaptive average pooling, identity mapping and Bilinear interpolation to W i ×W i The resolution of , where 1≤i,j≤M.

[0081] Subsequently, a 3×3 convolution is applied to smooth each resized feature map. The formula is where θ ij represents the parameters of smoothed convolution, is the j-th smoothed feature map of the i-th level.

[0082] After resizing all the i-th level feature maps to the same resolution, we apply the element-wise Hadamard product to all the resized feature maps to enhance the i-th level features with more semantic information and finer details as follows:

[0083] Where H(·) represents the Hadamard product.

[0084] after, is sent to the i-th level decoder for further resolution reconstruction and segmentation.

[0085] In some feasible embodiments, the working process of the DySample dynamic sampler is as follows: Figure 3 As shown, specifically:

[0086] Given an upsampling scale factor s and a feature map χ of size C×H×W, we further generate a point-by-point "dynamic range factor" by linearly projecting the input features. By using the sigmoid function and a static factor of 0.5, we generate a dynamic range factor of size 2gs. 2 ×H×W features, and then after pixel shuffling, we get an offset O of size 2g×sH×sW. The formula is as follows:

[0087] O=0.5sigmoid(linear1(χ))·linear2(χ)

[0088] Next, the original sampling network g of size 2g×sH×sW is superimposed with the offset O to obtain the sample S. That is, S=g+O.

[0089] Next, we use the grid sample function to resample the assumed bilinear interpolation χ using the positions in S to generate an upsampled feature map χ′ of size C×sH×sW. The process is defined as χ′=gridsample(χ,S).

[0090] In some feasible embodiments, when calculating the model loss, the MPDIoU loss function is used to further optimize the bounding box regression loss to detect more occluded targets, thereby ensuring that the target detection model maintains a high accuracy while also improving the computational efficiency.

[0091] The MPDIoU-based loss metric method also minimizes the loss function during the training phase to make the bounding box predicted by the model close to its true bounding box. Specifically, each predicted bounding box is:

[0092]

[0093] The ground-truth bounding box is approximated by minimizing the following loss function:

[0094] B gt =[x gt ,y gt ,w gt ,h gt ] T

[0095] L=min θ L(B gt ,B prd |θ)

[0096] Among them, B gt is the set of true bounding boxes, and θ is the parameter of the regression depth model. Finally, the loss function L based on MPDIoU is characterized by the following formula: MPDIoU .

[0097] L MPDIoU =1-MPDIoU

[0098] Therefore, all factors of the existing bounding box regression loss function can be determined by the four point coordinates. The conversion formula is as follows:

[0099]

[0100] Among them, |C| means covering B gt and B prd The minimum enclosed holding area, and Represents the coordinates of the center points of the real bounding box and the predicted bounding box respectively. gt and h gt represents the width and height of the ground-truth bounding box, w prd and h prd Represents the width and height of the predicted bounding box.

[0101] From the formula, we can see that all factors considered in the existing loss function, such as non-overlapping area, center point distance, width and height deviation, can be determined by the coordinates of the upper left point and the lower right point, which means that the L MPDIoU Not only is it thoughtful, but it also simplifies the calculation process.

[0102] If the aspect ratio of the predicted bounding box and the true bounding box is the same, then the predicted bounding box L is within the true bounding box MPDIoUThe value is lower than the predicted box outside the true bounding box. This feature ensures the accuracy of bounding box regression and makes the predicted bounding box less redundant. The MPDIoU algorithm takes into account the overlap area, center point distance, and width and height deviations, which can provide more accurate bounding box regression results when dealing with occlusion problems.

[0103] This means that even when the object is partially occluded, the MPDIoU algorithm can better predict the bounding box of the object, thereby improving the detection accuracy.

[0104] In some feasible embodiments, a refined strategy is adopted in the design of the detection head in order to achieve a balance between computational efficiency and detection accuracy.

[0105] Table 1 Reconstruction design of lightweight head detection

[0106] Original channel number Lightweight channel count Fifth ELAN module [192,192,96,3] [128,128,64,3] Sixth ELAN module [128,128,64,3] [96,96,48,3] Seventh ELAN module [192,192,96,3] [128,128,64,3] Eighth ELAN module [256,256,128,3] [192,192,96,3]

[0107] As can be seen from Table 1, by reducing the number of channels of the four ELAN modules, we have greatly reduced the overall parameters and calculation amount of the model and improved the detection efficiency.

[0108] In some feasible embodiments, during the model training process, a warm-up training method is used to gradually increase the learning rate from a smaller value until the set initial learning rate is reached. This can avoid the gradient being too large during initialization, making the trained model more stable.

[0109] Table 2 Training parameter configuration table

[0110] Optimizer SGD Batch size (Batch_size) 16 Epochs 300 Warmup epochs 5 Input image size (Input_size) 640 Learning rate 0.01 Warmup learning rate 1e-7

[0111] In some feasible embodiments, it also includes

[0112] The effectiveness of the infrared target detection model is verified.

[0113] In order to verify the effectiveness of the high-precision optimization method and SDI (Semantic and Detail Infusion) module, DySample dynamic upsampler, MPDIoU loss function and lightweight detection head network used in this experiment, this embodiment set up ablation experiments and comparative experiments to explore the impact of the proposed improvement method and some other improvement methods on the performance of the YOLOv9 model.

[0114] First of all, the performance indicators. In the YOLO series of models, the indicators for evaluating its network performance are mainly the following: Precision (P), Recall (R), and Mean Average Precision (mAP). This experiment uses mAP50 and mAP50:95 as performance reference indicators. mAP50 and mAP50:95 represent the mAP value when the IoU threshold is 0.5 and the average mAP value when the IoU starts from 50% and increases to 95% with a step size of 0.05. The larger the mean average precision mAP, the higher the overall accuracy of the model. The calculation formulas for each indicator are as follows:

[0115]

[0116] Among them, TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, and FP is the number of incorrectly predicted positive samples.

[0117] When discussing the balance between lightweight and model performance, the number of model parameters and the amount of computation must also be considered. Therefore, it is necessary to introduce two model architecture detail parameters: FLOPs (Floating Point Operations) and Params.

[0118] Method comparison

[0119] This embodiment adopts the method of ablation experiment and comparative experiment to partially verify the effectiveness of the improved algorithm. The ablation experiment is shown in Table 2. In the process of the ablation experiment, in order to verify the effectiveness of the SDI (Semantic and Detail Infusion) module, DySample dynamic upsampler, MPDIoU loss function and reconstructed lightweight detection head network, they are decomposed into multiple combinations for comparative tests.

[0120] First, YOLOv9s was used as the baseline algorithm (Baseline) to experiment on this dataset. After the experiment, it was found that YOLOv9s has room for improvement in accuracy, so subsequent improvements are also to improve and optimize it under the premise of not increasing the complexity of the model too much.

[0121] As shown in Table 3, based on the YOLOv9s baseline model, the performance of the model is reduced when the SDI (Semantic and Detail Infusion) module or the DySample dynamic upsampler is used alone, although SDI increases the model capacity, while DySample slightly reduces the performance without increasing the model complexity. However, when these two strategies are used in combination, the best object detection performance is achieved. Specifically, the YOLOv9s model combined with SDI and DySample improves the mAP50 to 53.4% ​​and the mAP50-95 to 33.1% on this dataset. At the same time, the number of parameters is 10.1M, the amount of calculation is 40.9G FLOPs, and only 280 epochs are trained, which is a comprehensive improvement compared with the baseline model. SDI enhances the expressiveness of the model by injecting semantic and detail features, while DySample improves the generalization performance of the model by dynamically adjusting the distribution of training samples. The combination of the two produces a synergistic effect, so this solution is effective. By reconstructing the detection head network, although the accuracy of the model has decreased, the number of parameters and the amount of calculation have dropped to 7.6M and 30.5FLOPs / G respectively, which greatly speeds up the detection rate. At the same time, by introducing the MPDIoU loss function, the model can better comprehensively consider the overlapping area, center point distance, and width and height deviations to improve the model accuracy. It can be seen that the mAP50% and mAP50-95% of the model have increased to 54.1% and 33.6% respectively, and the number of training rounds has also been greatly reduced to 240 rounds. It can be concluded that combining these four methods is effective and has a significant effect on improving the model.

[0122] Table 3. Comparison of ablation experiments

[0123]

[0124]

[0125] Under the premise of keeping the experimental environment unchanged, the performance of Light-SDM-YOLO is compared with the currently commonly used target detection algorithms: SSD, FasterRCNN, and YOLOv5s. The comparison results are shown in Table 4. By comparing the experimental results, SD-YOLOv9s shows excellent performance in multiple dimensions such as detection accuracy, robustness, model efficiency, and training speed, which fully proves the effectiveness and superiority of this method. Specifically, Light-SDM-YOLOv9s achieves 54.1% in the main evaluation index of mAP50%, which is significantly better than the traditional methods SSD, Faster-RCNN, and YOLOv5s. In the comprehensive evaluation index of mAP50-95%, its level of 33.6% is also much higher than other methods, indicating that our method not only has excellent performance on easy-to-detect targets, but also has strong robustness on difficult-to-detect targets. It is worth noting that Light-SDM-YOLOv9s only requires 7.6M parameters and 30.5G FLOPs of computation, which are lower than SSD and YOLOv5s in terms of parameters and computational complexity, which lays a solid foundation for practical applications. In addition, Light-SDM-YOLOv9s achieved the above excellent results in 240 epochs of training time, while SSD and YOLOv5s required 300 epochs, which highlights the efficiency of our method. Therefore, the SD-YOLO solution can well balance the lightweight model and algorithm performance, and can also outperform some common algorithms.

[0126] Table 4 Comparison of some target detection algorithms

[0127]

[0128] like Figure 4 As shown, a high-precision infrared target detection system based on improved YOLOv9 includes:

[0129] A training set building module, used for collecting self-coherent holograms and building a training set;

[0130] A network training module is used to construct an elimination network and perform training based on a training set to obtain a trained elimination network;

[0131] The elimination network includes a convolution layer, a normalization layer, an activation function and a deconvolution layer;

[0132] The output module is used to obtain the back-propagated hologram and input it into the trained elimination network to obtain a focused image.

[0133] In some embodiments, the training set construction module further includes:

[0134] A system building unit, used to build an imaging system to collect self-coherent holograms;

[0135] The phase shift unit, based on the phase shift algorithm, obtains the clearly focused sample information according to the self-coherent hologram and records the reconstruction distance;

[0136] A reverse propagation unit, used for performing reverse propagation at the same reconstruction distance to obtain a reverse propagation hologram;

[0137] A label setting unit, when there is a focused image in the back propagation distance, uses the information of the focused image as a label; when there is no focused image in the back propagation distance, uses a zero-value image as a label;

[0138] A construction unit is used to construct a training set based on the back-propagated holograms and the corresponding labels.

[0139] In some embodiments, it also includes:

[0140] A data expansion unit, used for superimposing the self-coherent holograms to obtain a superimposed hologram;

[0141] In some embodiments, the network training module includes:

[0142] A network construction unit, used to construct an elimination network;

[0143] The error function setting unit uses the root mean square error as the loss function of the network training, and trains the elimination network according to the self-coherent holograms and corresponding labels in the training set;

[0144] The parameter adjustment unit is used to adjust the parameters of the elimination network according to the convergence of the loss function to obtain the trained elimination network.

[0145] In some embodiments, it also includes:

[0146] The testing unit is used to construct a test set based on the back-propagation hologram and the corresponding labels; and to test the trained elimination network based on the test set.

[0147] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0148] A high-precision infrared target detection device based on YOLOv9 improvement:

[0149] at least one processor;

[0150] at least one memory for storing at least one program;

[0151] When the at least one program is executed by the at least one processor, the at least one processor implements the high-precision infrared target detection method based on the improvement of YOLOv9 as described above.

[0152] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0153] A storage medium stores processor-executable instructions, which are used to implement a high-precision infrared target detection method based on an improvement of YOLOv9 when executed by a processor.

[0154] The contents of the above method embodiments are all applicable to the present storage medium embodiments. The functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0155] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A high-precision infrared target detection method based on improved YOLOv9, characterized in that: The following steps are involved: Obtain image data and build a training set; Based on the YOLOv9 basic framework, the SDI module and DySample dynamic sampler are introduced to build an infrared target detection model; Training the infrared target detection model based on the training set to obtain a trained infrared target detection model; The image to be tested is input into the trained infrared target detection model, and the infrared target detection result is output.

2. According to claim 1, a high-precision infrared target detection method based on YOLOv9 improvement is characterized in that: Also includes: Data enhancement is performed on the image data based on the illumination intensity and the rotation direction.

3. According to claim 1, a high-precision infrared target detection method based on YOLOv9 improvement is characterized in that: The infrared target detection model includes a backbone network and a detection head, wherein: The backbone network includes a convolutional layer, a pooling layer and an ELAN module; The detection head includes an upsampling merging module, a downsampling merging module, an ELAN block and an output convolutional layer; Between the two downsampling merging modules, an SDI module is embedded to replace the Concat operation; In the upsampling merging module, a DySample dynamic sampler is introduced.

4. According to claim 3, a high-precision infrared target detection method based on YOLOv9 improvement is characterized in that: The step of inputting the image to be tested into the trained infrared target detection model and outputting the infrared target detection result specifically includes: Inputting the image to be tested into the trained infrared target detection model; Processing the image to be tested based on the backbone network, reducing the input resolution, extracting semantic features and enhancing target feature representation, to obtain first-stage features; Based on the detection head processing the features of the first stage, the semantic features at different levels are integrated through sampling and feature fusion to obtain detection features; An infrared target detection result is output based on the detection feature.

5. According to claim 3, a high-precision infrared target detection method based on YOLOv9 improvement is characterized in that: The working principle of the SDI module includes: Through the upsampling process, the size of the feature is increased and matched to the size of the current level; Through the downsampling process, the size of the features is reduced and matched to the size of the corresponding level; Spatial and channel attention mechanisms are applied to the features of each level generated by the encoder.

6. According to claim 1, a high-precision infrared target detection method based on YOLOv9 improvement is characterized in that: Also includes: During the training process of the infrared target detection model, the MPDIoU loss function is used to optimize the bounding box regression loss.

7. A high-precision infrared target detection system based on YOLOv9 improvement, characterized in that: include: A data acquisition unit, used to acquire image data and construct a training set; The model building unit, based on the YOLOv9 basic framework, introduces the SDI module and DySample dynamic sampler to build an infrared target detection model; A model training unit, which trains the infrared target detection model based on the training set to obtain a trained infrared target detection model; The detection unit is used to input the image to be detected into the trained infrared target detection model and output the infrared target detection result.

8. A high-precision infrared target detection device based on YOLOv9 improvement, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a high-precision infrared target detection method based on an improvement of YOLOv9 as described in any one of claims 1-6.

Images