An infrared unmanned aerial vehicle target detection algorithm based on deep learning
By using the ScYOLO neural network and Mosaic data augmentation algorithm, the challenges of infrared image detection performance degradation and embedded device deployment were solved, enabling efficient and real-time infrared UAV target detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2023-07-10
- Publication Date
- 2026-05-29
AI Technical Summary
When existing deep learning target detection algorithms are applied to infrared images, their detection performance degrades, making it difficult to meet real-time requirements and unsuitable for deployment on embedded AI devices. Traditional infrared target detection algorithms suffer from high computational complexity, poor robustness, and weak generalization ability.
The ScYOLO neural network, which uses a single-channel input and includes a backbone network G-ELAN, a neck network, and a head network, is combined with the Mosaic data augmentation algorithm and a lightweight design to perform infrared UAV target detection through offline training and online inference.
It improves the speed and accuracy of infrared image detection, is suitable for deployment in embedded AI devices, has good robustness and generalization ability, and meets real-time requirements.
Smart Images
Figure CN117011725B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an infrared unmanned aerial vehicle (UAV) target detection algorithm, belonging to the field of image processing technology. Background Technology
[0002] With the increasing use of drones, while bringing convenience, they also pose significant security risks in both military and civilian fields, threatening national defense, public safety, and privacy. Therefore, effectively detecting, monitoring, and regulating the use of drones is an important issue that urgently needs to be addressed.
[0003] Target detection technologies for unmanned aerial vehicles (UAVs) mainly include radar detection, wireless detection, audio detection, visible light detection, and infrared detection. Current UAVs mostly employ stealth and miniaturization designs and rely on electric motors for low-altitude, slow flight, generating relatively little noise. Therefore, radar and audio detection technologies are less effective. Visible light detection is easily affected by weather and lighting conditions, especially at night when detection is impossible. Wireless detection also fails when the UAV maintains radio silence. Infrared target detection, on the other hand, uses infrared sensors to image targets based on differences in infrared thermal radiation. It is unaffected by extreme weather and lighting conditions, allowing for all-weather target detection. It boasts advantages such as wide detection range, long detection distance, high accuracy, strong penetration ability, and strong anti-interference capabilities, making infrared target detection technology a focus of attention across various fields.
[0004] Traditional infrared target detection algorithms suffer from the following problems: First, manually designing and selecting effective image features requires a large amount of prior knowledge, resulting in low overall development efficiency. Second, redundant windows are generated during candidate region selection, leading to excessively high computational complexity, slow detection speed, and potential impact on subsequent classification performance, thus affecting the algorithm's real-time performance and accuracy. Third, manually designed image features are typically tailored to specific detection tasks. If conditions such as the target, application background, detection distance, and angle change, the detection performance will be significantly reduced, exhibiting poor robustness to environmental changes, weak generalization ability, and limited detection accuracy under complex background conditions.
[0005] Deep learning-based object detection algorithms can autonomously learn feature representations in image data, offering advantages such as fast detection speed and strong generalization ability, and have gradually replaced traditional object detection algorithms. However, existing deep learning object detection algorithms still face the following problems when applied to infrared images: First, most are based on visible light RGB images, extracting high-level semantic features from image color, target shape, and detailed texture information. However, infrared images lack color and detailed texture, and are characterized by uneven brightness, blurred details, and poor contrast, containing less feature information, which affects the feature representation of the network. Directly transferring visible light object detection networks to infrared scenes will result in decreased detection performance. Therefore, targeted optimization of network structure design and image preprocessing is required. Second, most object detection networks applied directly to infrared images have limited accuracy, and the number of parameters and computational load are too large. They have high hardware requirements, making it difficult to meet real-time requirements and unsuitable for deployment on embedded AI devices. Therefore, lightweight design and quantization are needed. Summary of the Invention
[0006] This invention addresses the problem that existing deep learning target detection algorithms suffer from performance degradation when applied to infrared images, making it difficult to meet real-time requirements and unsuitable for deployment on embedded AI devices. Therefore, it proposes a deep learning-based infrared drone target detection algorithm.
[0007] The technical solution adopted by this invention to solve the above problems is as follows: The infrared UAV target detection algorithm based on deep learning described in this invention is implemented through a single-channel input neural network ScYOLO, which consists of a backbone network G-ELAN, a neck network, and a head network; the infrared UAV target detection algorithm based on deep learning is implemented through the following steps:
[0008] Step 1: Perform offline training using the ScYOLO neural network;
[0009] Step 2: Perform online inference using the ScYOLO neural network.
[0010] Furthermore, in the offline training phase of step one, the Mosaic data augmentation algorithm is introduced. Four images are randomly selected, randomly transformed, and then synthesized into a new image. The new image contains multiple drone targets and four different backgrounds. The specific process of the Mosaic data augmentation algorithm is as follows:
[0011] Step 1: Read a batch from the training dataset and randomly select four images from it;
[0012] Step 2: Randomly flip, scale, and distort the length and width of each of the four images.
[0013] Step 3: In the central region of the image, randomly set the center cropping position (x, y), and according to the set output size (w×h), restrict x∈[0.3w,0.7w], y∈[0.3h,0.7h];
[0014] Step 4: Crop the images. Crop the four images one by one using the crosshairs at the cropping positions. The top left image is cropped to the top left part of the cropping position, the top right image is cropped to the top right part of the cropping position, and the bottom left and bottom right images are cropped in the same way. At the same time, the target box of the original image needs to be adjusted so that it does not exceed the cropping range of the original image.
[0015] Step 5: Stitch the images together to obtain a new image;
[0016] Step 6: Repeat the above process BatchSize times to obtain BatchSize stitched new images, which are then fed into the neural network for training.
[0017] Furthermore, before using the ScYOLO neural network for online inference in step two, ten warm-ups are performed, and the batch size is set to 1 during inference. The model's inference time includes data processing time and network feedforward computation time. The model's inference steps are as follows:
[0018] Step 2 (a) Image preprocessing: Adjust the size of the input image using bicubic interpolation to uniformly adjust it to 416×416, and normalize the image data.
[0019] Step 2 (ii): Feedforward calculation of the network. The preprocessed image data is fed into the neural network for feedforward calculation.
[0020] Step 2 (3) Data post-processing: Decode the network output and perform nonmaximum suppression to obtain the detection result;
[0021] Step 2 (iv): Draw the detection results. Draw the network prediction box, target category, and confidence score on the original input image.
[0022] Furthermore, the backbone network G-ELAN sets the number of input image channels to 1, the input size to 416×416, and the maximum width of the entire network to 256. G-ELAN mainly consists of a Focus module, a multi-branch stacked module, and a max pooling layer. The Focus network is placed at the front of the backbone network. It samples values from one channel of the image at row and column intervals, expanding one channel into four channels and converting the feature information in the width and height dimensions into the channel dimension. For a single-channel input image, the Focus network expands it into four channels.
[0023] Furthermore, the Ghost module is introduced into the backbone network G-ELAN for model lightweighting. The steps to improve the Ghost module are as follows:
[0024] Step a: Replace the ReLU activation function with the Hardswish activation function;
[0025] Step b: The kernel size of the first convolutional layer in the module is adjusted from 1×1 to 3×3, and the kernel size of the second grouped convolutional layer is adjusted from 3×3 to 5×5. This expands the receptive field of the convolutional layers, resulting in a more significant performance improvement at the cost of slightly increasing the amount of computation and parameters.
[0026] Furthermore, the neck network is a combination of FPN and PAN, and uses the MCB module for enhanced feature extraction. In the neck network, a 3×3 reparameterized convolution RepConv with a stride of 2 is used for downsampling of the feature maps in the PAN structure.
[0027] Furthermore, the head network employs a decoupled detection head and depthwise separable convolutions.
[0028] The beneficial effects of this invention are:
[0029] 1. This invention is based on a deep learning algorithm. Compared with traditional target detection algorithms, it does not require the design of effective infrared image features by humans. ScYOLO can learn the feature representation in image data autonomously. The algorithm has stronger robustness and generalization ability, and does not require too much prior image knowledge, which can improve project development efficiency.
[0030] 2. This invention is specifically designed based on the characteristics of infrared images and the target characteristics of UAVs. For infrared images, compared with ordinary target detection networks, it can achieve a better balance between speed and accuracy with a small number of parameters and computational load.
[0031] 3. This invention is specifically optimized for hardware deployment, using activation functions suitable for hardware, with a lightweight network structure and merged network layers. Compared to ordinary object detection networks, it reduces complex branch structures and redundant layers, making it suitable for various hardware frameworks and more suitable for embedded hardware deployment. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the ScYOLO neural network structure;
[0033] Figure 2 This is a schematic diagram of the Ghost convolutional module;
[0034] Figure 3 This is a structural diagram of the MCB module;
[0035] Figure 4 This is a diagram illustrating the Mosaic cutting and splicing method;
[0036] Figure 5 This is a schematic diagram of the Mosaic stitched image and detection results;
[0037] Figure 6 This is a schematic diagram of the ScYOLO training error variation curve;
[0038] Figure 7 This is a schematic diagram of the ScYOLO single-frame detection results. Detailed Implementation
[0039] Specific implementation method one: Combining Figures 1 to 7 This embodiment describes a deep learning-based infrared UAV target detection algorithm. This algorithm is implemented using a single-channel input neural network, ScYOLO, which consists of a backbone network G-ELAN, a neck network, and a head network. The deep learning-based infrared UAV target detection algorithm is implemented through the following steps:
[0040] Step 1: Perform offline training using the ScYOLO neural network;
[0041] Step 2: Perform online inference using the ScYOLO neural network.
[0042] At the input end during inference and prediction, the infrared image does not need to be converted into a three-channel image or undergo other image preprocessing before being input into the neural network; only [the following process is required]. Normalization allows the data to be fed into the neural network for prediction, greatly improving the algorithm's running speed. In the above formula, I... Norm I represents the normalized image data, and I represents the original input image data.
[0043] During offline training, conventional data augmentation techniques for neural networks were used at the input end, such as random horizontal flipping and multi-scale random scaling. In addition, to address the uneven brightness of infrared single-channel grayscale images, grayscale jitter was applied to the image, i.e., random linear grayscale changes.
[0044] The dataset used in this invention contains mostly infrared images with only one target, and only a few images with two targets simultaneously. Considering that real-world applications may require detecting multiple different drone targets with varying shapes, orientations, and angles, this paper introduces the Mosaic data augmentation algorithm during the training phase to improve the network's multi-target detection capability, enrich the targets and backgrounds in the training samples, and further enhance the network's generalization ability and robustness. Four randomly selected images are randomly transformed and synthesized into a new image containing multiple drone targets and four different backgrounds. This new image, fed into the neural network, significantly enriches the small targets and detection backgrounds, and simultaneously calculates the activation statistics of the four images during Batch Normalization (BN) layer computation.
[0045] Specific Implementation Method Two: Combining Figures 1 to 7 This embodiment describes a deep learning-based infrared UAV target detection algorithm. In the offline training phase, the Mosaic data augmentation algorithm is introduced. Four images are randomly selected, randomly transformed, and then synthesized into a new image. This new image contains multiple UAV targets and four different backgrounds. The specific process of the Mosaic data augmentation algorithm is as follows:
[0046] Step 1: Read a batch from the training dataset and randomly select four images from it;
[0047] Step 2: Randomly flip, scale, and distort the length and width of each of the four images.
[0048] Step 3: In the central region of the image, randomly set the center cropping position (x, y), and according to the set output size (w×h), restrict x∈[0.3w,0.7w], y∈[0.3h,0.7h];
[0049] Step 4: Crop the images. Crop the four images one by one using the crosshairs at the cropping positions. The top left image is cropped to the top left part of the cropping position, the top right image is cropped to the top right part of the cropping position, and the bottom left and bottom right images are cropped in the same way. At the same time, the target box of the original image needs to be adjusted so that it does not exceed the cropping range of the original image.
[0050] Step 5: Stitch the images together to obtain a new image;
[0051] Step 6: Repeat the above process BatchSize times to obtain BatchSize stitched new images, which are then fed into the neural network for training.
[0052] The final result of the cutting and piecing process described above is shown in the attached image. Figure 5As shown. Mosaic image stitching enhancement is turned off before the end of training because the mosaic images deviate significantly from the distribution of real infrared images. The appropriate enhancement strategy varies depending on the model size. This invention uses a lightweight, small model, and for small models, data augmentation should be moderately reduced. Therefore, during training, Mosaic enhancement is only enabled in the first 70% of epochs. In this paper, all neural networks are trained for 400 epochs, meaning that Mosaic data augmentation is enabled in the first 280 epochs.
[0053] Comparative studies have shown that performing complex traditional image processing, such as guided filtering, CLAHE contrast enhancement, and gamma transform, before inputting images into the neural network can improve the visual discrimination of targets, but it has very limited effect on improving the prediction accuracy of the neural network and greatly increases the computational load, reducing real-time performance. Therefore, during training, only a simple normalization operation is performed on the augmented infrared images, calculating the mean and standard deviation of the entire infrared dataset, and normalizing each image, such as... As shown.
[0054] The loss function of ScYOLO consists of three parts: regression loss (Reg Loss), confidence loss (Obj Loss), and classification loss (Cls Loss). The regression loss for the bounding box position uses GIoU Loss, the classification loss uses BCE Loss, and the confidence loss uses Varifocal Loss.
[0055] The network training in this invention does not use pre-trained weights. It is trained from scratch on a self-built dataset for 400 epochs using a stochastic gradient descent (SGD) optimizer with SGD weight decay set to 0.0005 and momentum set to 0.937. FP16 mixed precision training is used. During ScYOLO training, the batch size is set to 64, the maximum and minimum learning rates are set to 0.01 and 0.0001 respectively, and the initial learning rate lr = 0.01 is adjusted according to lr × BatchSize / 64. Cosine annealing is used to decay the learning rate to avoid getting trapped in local optima.
[0056] The error variation during ScYOLO training is shown in the attached figure. Figure 6 As shown, the error curve eventually becomes constant during training, proving that it can converge normally.
[0057] Specific implementation method three: Combining Figures 1 to 7This embodiment describes a deep learning-based infrared UAV target detection algorithm. Before online inference using the ScYOLO neural network in step two, ten warm-ups are performed, and the batch size is set to 1 during inference. The model's inference time includes data processing time and network feedforward computation time. The model's inference steps are as follows:
[0058] Step 2 (a) Image preprocessing: Adjust the size of the input image using bicubic interpolation to uniformly adjust it to 416×416, and normalize the image data.
[0059] Step 2 (ii): Feedforward calculation of the network. The preprocessed image data is fed into the neural network for feedforward calculation.
[0060] Step 2 (3) Data post-processing: Decode the network output and perform nonmaximum suppression to obtain the detection result;
[0061] Step 2 (iv): Draw the detection results. Draw the network prediction box, target category, and confidence score on the original input image.
[0062] Specific implementation method four: Combination Figures 1 to 7 This embodiment describes a deep learning-based infrared UAV target detection algorithm with a backbone network called G-ELAN. The input image has one channel and an input size of 416×416. The maximum width of the entire network is 256. G-ELAN mainly consists of a Focus module, a multi-branch stacking module, and a max-pooling layer. The Focus network is placed at the front of the backbone network. It samples values from one channel of the image at row and column intervals, expanding one channel into four channels. It converts the feature information in the width and height dimensions into the channel dimension. For a single-channel input image, the Focus network expands to four channels.
[0063] The Focus network, placed at the front of the backbone network, samples values from one channel of the image at row and column intervals, expanding one channel into four channels. It transforms the feature information in the width and height dimensions into channel dimensions. For a single-channel input image, the Focus network can expand it into a four-channel network. For infrared images, standard convolution generates a large number of redundant feature maps, some of which may be similar to each other. This not only fails to significantly improve the detection performance of infrared targets but also increases the number of network parameters and computational cost. This is based on the characteristics of infrared images and the characteristics of UAV targets.
[0064] Specific Implementation Method Five: Combining Figures 1 to 7This embodiment describes a deep learning-based infrared UAV target detection algorithm that introduces a Ghost module into the backbone network G-ELAN for model lightweighting. The steps for improving the Ghost module are as follows:
[0065] Step a: Replace the ReLU activation function with the Hardswish activation function;
[0066] Step b: The kernel size of the first convolutional layer in the module is adjusted from 1×1 to 3×3, and the kernel size of the second grouped convolutional layer is adjusted from 3×3 to 5×5. This expands the receptive field of the convolutional layers, resulting in a more significant performance improvement at the cost of slightly increasing the amount of computation and parameters.
[0067] Specific Implementation Method Six: Combination Figures 1 to 7 This embodiment describes a deep learning-based infrared UAV target detection algorithm whose neck network is a combination of FPN and PAN structures. It uses an MCB module for enhanced feature extraction and employs a 3×3 reparameterized convolution RepConv with a stride of 2 in the neck network for downsampling of the feature maps in the PAN structure.
[0068] like Figure 1 As shown, multi-scale feature representations are generated to detect and identify targets at various scales, expanding the detection range and further improving detection accuracy. A 3×3 reparameterized convolution RepConv with a stride of 2 is also used in the neck network for downsampling of feature maps in the PAN structure to enhance feature representation and further improve the feature pyramid feature fusion effect.
[0069] The MCB module is used to extract image features. It is designed for infrared images based on ELAN. Without disrupting the original gradient path or affecting parameter utilization, it uses the improved Ghost convolution module mentioned above to replace some ordinary convolutions for lightweighting. Its structure is shown in the attached figure. Figure 3 As shown, G-ELAN is mainly composed of stacked MCB modules. At the structural level, G-ELAN only changes the structure of the computing modules, while the structure of the transition layer remains unchanged.
[0070] Specific implementation method seven: Combination Figures 1 to 7 This embodiment describes a deep learning-based infrared UAV target detection algorithm whose head network employs a decoupled detection head and depthwise separable convolutions.
[0071] Decoupling the detection head not only improves accuracy but also accelerates network convergence. ScYOLO uses the Hardswish activation function, which is more suitable for hardware deployment. Hardswish deployment on embedded mobile devices offers several advantages: First, the Hardswish activation function is primarily based on ReLU6 (with a maximum output limit of 6), and ReLU6 has been optimized for implementation on most hardware and software frameworks. Second, it prevents accuracy degradation caused by different implementations of the Sigmoid function during quantization. Finally, it can be implemented as a piecewise function, reducing frequent memory accesses and improving inference efficiency. The ReLU activation function is only used in the downsampling RepConv module of the Neck section.
[0072] To further reduce the number of parameters and network layers and improve inference speed, ScYOLO's convolutional layers and Batch Norm layers were fused into a single biased convolutional layer during the test inference phase. After merging the network layers, the entire neural network consists of only three basic structures: convolutional layers, max pooling layers, and upsampling layers, which facilitates subsequent hardware implementation.
[0073] At the output end, an anchor-free detection mechanism is used. This mechanism eliminates the need for prior clustering analysis, significantly reducing the number of hyperparameters required and avoiding many anchor design techniques, making the entire network simpler and clearer, especially greatly simplifying the network training and post-processing decoding stages. The anchor-free detection head incorporates downsampling size information from the backbone network during training and decoding. Each grid in the image has only one prediction box, and each prediction box directly predicts the height, width, and coordinate offset of the top-left corner of the grid. This reduces the number of network parameters and computational cost, accelerates network convergence, and greatly reduces inference latency.
[0074] During online reasoning:
[0075] The single-frame detection results of ScYOLO on the server are attached. Figure 7 As shown, the red box represents a rotary-wing drone, the blue box represents a fixed-wing drone, and the green box represents a real frame. ScYOLO can accurately detect drone targets of various sizes, backgrounds, and whether they are occluded or blurred, and can perform real-time monitoring of multiple targets.
[0076] The comparison test results of ScYOLO with SSD, Faster R-CNN, YOLOv3, YOLOv4, YOLOv5, YOLOX, and YOLOv7 are shown in Appendix Table 1.
[0077] Appendix 1: Comparison Test Results of Various Networks
[0078]
[0079] All networks were tested with identical parameter settings, primarily using mAP. 0.5 The FPS metric measures the model's accuracy and speed. For mAP calculation, the confidence threshold for ScYOLO and Faster R-CNN was set to 0.01, while others were set to 0.001. The NMS threshold was set to 0.5, and the BatchSize was set to 1.
[0080] The ScYOLO invention can be quantized and deployed on Jetson AGX Xavier. The evaluation results under the optimal operating mode (maximum power consumption) of Jetson AGX Xavier are shown in Appendix Table 2. The optimal operating mode requires setting AGX Xavier to operate in mode 0 (MAXN), starting all hardware resources, unlocking its maximum operating frequency, performing 10 warm-ups before actual inference, and setting the BatchSize to 1 during inference.
[0081] The model's inference time includes data processing time and network feedforward computation time. Appendix Table 3 shows the average processing time and average network computation time of the four networks deployed with FP16 accuracy. Compared to other networks, ScYOLO significantly reduces network data processing time and substantially improves FPS.
[0082] After TensorRT FP16 deployment on each network, the FPS under the 10W, 15W, 30W, and MAXN optimal operating modes are shown in Appendix Table 4. ScYOLO can still achieve 24 FPS in the lowest power mode of 10W, which basically meets the real-time requirements. It is higher than 25 FPS in both 15W and 30W modes, and can even reach 67 FPS at the highest power mode.
[0083] Appendix 2: Evaluation Results of YOLOX-Nano and ScYOLO
[0084]
[0085] Appendix 3: Data Processing and Network Computing Time
[0086]
[0087] Appendix 4 Speed under different power consumption modes
[0088]
[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A deep learning-based infrared UAV target detection algorithm, characterized in that: The deep learning-based infrared UAV target detection algorithm is implemented using a single-channel input neural network ScYOLO, which consists of a backbone network G-ELAN, a neck network, and a head network. The deep learning-based infrared UAV target detection algorithm is implemented through the following steps: Step 1: Perform offline training using the ScYOLO neural network; Infrared images are processed before being input into the neural network. Normalization During the offline training phase, the Mosaic data augmentation algorithm is introduced. Four images are randomly selected, randomly transformed, and then synthesized into a new image. The new image contains multiple drone targets and four different backgrounds. The specific process of the Mosaic data augmentation algorithm is as follows: Step 1: Read a batch from the training dataset and randomly select four images from it; Step 2: Randomly flip, scale, and distort the length and width of each of the four images. Step 3: Randomly set the center cropping position in the center area of the image. According to the set output size ,limit ; Step 4: Crop the images. Crop the four images one by one using the crosshairs at the cropping positions. The top left image is cropped to the top left part of the cropping position, the top right image is cropped to the top right part of the cropping position, and the bottom left and bottom right images are cropped in the same way. At the same time, the target box of the original image needs to be adjusted so that it does not exceed the cropping range of the original image. Step 5: Stitch the images together to obtain a new image; Step 6: Repeat the above process BatchSize times to obtain BatchSize stitched new images, which are then fed into the neural network for training. During training, normalization is performed only on the augmented infrared images. The mean and standard deviation of the entire infrared dataset are calculated, and each image is normalized. The loss function of ScYOLO consists of three parts: regression loss, confidence loss, and classification loss. The regression loss for the bounding box location uses GIoU Loss, the classification loss uses BCE Loss, and the confidence loss uses Varifocal Loss. We trained from scratch on a self-built dataset for 400 epochs using a stochastic gradient descent optimizer with SGD weight decay set to 0.0005 and momentum set to 0.937, and FP16 mixed precision training. During ScYOLO training, the batch size was set to 64, the maximum and minimum learning rates were set to 0.01 and 0.0001 respectively, the initial learning rate lr=0.01, and it was adjusted according to lr×BatchSize / 64. Cosine annealing was used to decay the learning rate to avoid getting trapped in local optima. Step 2: Perform online inference using the ScYOLO neural network; Ten warm-up cycles were performed, with the batch size set to 1 during inference. The model's inference time includes data processing time and network feedforward computation time. The model's inference steps are as follows: Step 2 (a) Image preprocessing: Adjust the size of the input image using bicubic interpolation to uniformly adjust it to 416×416, and perform image data normalization; Step 2 (II): Feedforward calculation of the network. The preprocessed image data is fed into the neural network for feedforward calculation. Step 2 (3): Data post-processing, decoding the network output and performing nonmaximum suppression to finally obtain the detection result; Step 2 (iv): Draw the detection results. Draw the network prediction box, target category, and confidence score on the original input image. The backbone network G-ELAN is set to have 1 input image channel and an input size of 416×416. The maximum width of the entire network is 256. G-ELAN mainly consists of a Focus module, a multi-branch stacked module, and a max pooling layer. The Focus network is placed at the front of the backbone network. It samples values from one channel of the image at row and column intervals, expanding one channel into four channels. It transforms the feature information in the width and height dimensions into the channel dimension. For a single-channel input image, the Focus network expands it into four channels. At the output end, an anchor-free detection mechanism is used. The anchor-free detection head introduces downsampling size information from the backbone network during training and decoding. Each grid in the image has only one prediction box, and each prediction box directly predicts the height and width of the prediction box and the coordinate offset of the top left corner of the grid. The neck network is a combination of FPN and PAN, and uses the MCB module for enhanced feature extraction. The neck network uses a 3×3 reparameterized convolution RepConv with a stride of 2 for downsampling of the feature maps in the PAN structure.
2. The infrared UAV target detection algorithm based on deep learning according to claim 1, characterized in that: In the backbone network G-ELAN, the Ghost module is introduced to reduce model weight. The steps to improve the Ghost module are as follows: Step a: Replace the ReLU activation function with the Hardswish activation function; Step b: The kernel size of the first convolutional layer in the module is adjusted from 1×1 to 3×3, and the kernel size of the second grouped convolutional layer is adjusted from 3×3 to 5×5. This expands the receptive field of the convolutional layers, resulting in a more significant performance improvement at the cost of slightly increasing the amount of computation and parameters.
3. The infrared UAV target detection algorithm based on deep learning according to claim 1, characterized in that: The head network employs a decoupled detection head and depthwise separable convolutions.