A progressive unmanned aerial vehicle detection method and system based on super-resolution guidance

Abstract

The application discloses a kind of progressive unmanned aerial vehicle detection method and system based on super-resolution guide.Therein, to solve the problems of low detection accuracy of low-resolution unmanned aerial vehicle image, insufficient feature coordination, pseudo-texture interference and calculation redundancy in traditional "super-resolution + detection" series framework, a dual-resolution feature coordination architecture is proposed, which extracts features in parallel through fixed scale and random scale branches, and uses random scale features to guide fixed scale feature enhancement;It realizes target area focusing by using candidate box mapping mechanism, only the target area is enhanced by feature-level local super-resolution, avoiding invalid calculation of full-image super-resolution;Through cross-scale mapping loss and cosine similarity loss constraint, the spatial consistency and semantic alignment of the target area of the dual branch are ensured.The application realizes the deep coupling of super-resolution reconstruction and unmanned aerial vehicle detection, improves the low-resolution unmanned aerial vehicle target recognition, effectively suppresses the pseudo-texture interference, and improves the detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and more specifically, to a progressive UAV detection method and system based on super-resolution guidance. Background Technology

[0002] Drones, with their advantages of maneuverability, flexibility, and wide coverage, are widely used in various fields, and their monitoring largely relies on ground-based monitors. In visual monitoring tasks using ground-based monitors, the accuracy of the monitoring directly determines the effectiveness of the oversight. However, when ground-based monitors monitor drones in flight, due to objective factors such as long monitoring distances and offset shooting angles, the acquired drone images often suffer from low resolution and a lack of drone target feature information, posing a significant challenge to subsequent drone detection.

[0003] To address the aforementioned challenge of detecting low-resolution UAV images, existing technologies often employ a framework that combines super-resolution reconstruction with target detection. This involves first enhancing the image resolution using a super-resolution module, and then inputting the processed image into a detection module to identify the UAV.

[0004] However, this type of concatenated framework has obvious drawbacks: on the one hand, it ignores the essential difference between the super-resolution reconstruction target and the UAV detection target. The reconstruction target focuses on the fidelity of the image at the pixel level, while the detection target focuses on the semantic-level feature extraction of the UAV. The misalignment of the two targets leads to poor synergy, and the concatenated structure has a lot of redundant computation, resulting in extremely low overall computational efficiency. On the other hand, the existing super-resolution module only uses the image reconstruction quality as the single training target. The texture details it generates may not be suitable for the UAV recognition requirements, and may even introduce fake textures and additional noise. These invalid or interfering information will seriously interfere with the discrimination logic of the subsequent detection module, further reducing the detection accuracy of the UAV.

[0005] Therefore, how to overcome the limitations of the existing concatenated framework and achieve efficient collaboration between super-resolution reconstruction and UAV detection, while improving the recognition of low-resolution UAV target features, avoiding pseudo-texture interference and improving computational efficiency, has become a key issue that urgently needs to be addressed in the current scenario of ground monitors monitoring UAVs. Summary of the Invention

[0006] This invention provides a progressive UAV detection method and system based on super-resolution guidance to solve the problem of low target detection accuracy caused by the inefficient coordination between super-resolution reconstruction and UAV detection in the prior art.

[0007] To achieve the above objectives, this invention provides a progressive UAV detection method based on super-resolution guidance. The method includes: S1, acquiring original UAV images, processing them using both fixed and random resolutions, and inputting them into the backbone network of a detection model to extract fixed-scale and random-scale feature maps; S2, performing scale alignment and feature interaction between the random-scale and fixed-scale feature maps to generate an enhanced fixed-scale feature map; S3, predicting fixed-scale candidate boxes and random-scale candidate boxes based on the enhanced fixed-scale and random-scale feature maps, mapping the fixed-scale candidate boxes to the random-scale feature map, and determining the target based on the mapping results and the random-scale candidate boxes. S4. Based on the mapped candidate boxes and random-scale candidate boxes, crop local feature maps from the fixed-scale feature map and the aligned random-scale feature map, and perform super-resolution enhancement to obtain two super-resolution local feature maps; S5. Calculate the semantic consistency loss value of the fixed-scale candidate boxes and the two super-resolution local feature maps according to the cosine similarity loss function, and update the parameters of the detection model in combination with the candidate box mapping loss value; S6. Repeat S1~S5 until all loss values ​​and values ​​converge to obtain the final detection model; S7. Obtain the image of the drone to be detected, process it only with a fixed resolution, and input it into the final detection model for detection, outputting the drone category and candidate box coordinates.

[0008] Optionally, S2 includes: aligning the random-scale feature map with the fixed-scale feature map in the feature fusion module of the detection model to obtain an aligned feature map; performing element-wise multiplication of the aligned feature map with the fixed-scale feature map to obtain a modulation feature map; performing element-wise superposition of the modulation feature map with the fixed-scale feature map to obtain a preliminary enhanced feature map; and performing ReLU activation function operation on the preliminary enhanced feature map and batch normalization processing to obtain an enhanced fixed-scale feature map.

[0009] Optionally, step S3 includes: predicting target candidate boxes and their coordinates and confidence scores using the enhanced fixed-scale feature map and random-scale feature map respectively through the coarse detection head of the detection model; filtering the target candidate boxes according to a preset confidence threshold to remove background candidate boxes with confidence scores below the threshold, thereby obtaining fixed-scale candidate boxes and random-scale candidate boxes; calculating a scale conversion factor based on the ratio of fixed resolution to random resolution, and mapping the fixed-scale candidate boxes to the random-scale feature map according to the scale conversion factor to obtain mapped candidate boxes; performing boundary verification on the mapped candidate boxes and removing mapped candidate boxes that exceed the range of the random-scale feature map; calculating the intersection-union ratio (IU) of the filtered mapped candidate boxes and the random-scale candidate boxes, and retaining mapped candidate boxes with IU greater than or equal to a preset IU threshold; and calculating the candidate box mapping loss value based on the retained mapped candidate boxes and random-scale candidate boxes.

[0010] Optionally, the candidate box mapping loss value is calculated according to the following formula:

[0011] Lroi = α×L_reg+β×L_iou;

[0012] When |GT-Pred|≤1, L_reg=0.5×(GT-Pred)²; when |GT-Pred|>1, L_reg=|GT-Pred|-0.5;

[0013] L_iou = -ln(IOU);

[0014] Wherein, α is the first preset weight coefficient, β is the second preset weight coefficient, L_reg is the first coordinate regression loss value, L_iou is the intersection-union ratio loss value, Lroi is the candidate box mapping loss value, GT is the coordinates of the retained mapped candidate box, Pred is the coordinates of the random scale candidate box; and IOU is the mean of the intersection-union ratio of the retained mapped candidate box and the random scale candidate box.

[0015] Optionally, S4 includes: cropping corresponding local feature maps from the fixed-scale feature map and the random-scale feature map after alignment with the fixed-scale feature map, based on the coordinates of the retained mapped candidate boxes and the random-scale candidate boxes; inputting the cropped local feature maps into a local super-resolution module for feature-level super-resolution enhancement to obtain mapped local enhanced feature maps and random-scale local enhanced feature maps; the local super-resolution module includes a channel attention submodule and a residual learning submodule; performing noise filtering on the mapped local enhanced feature maps and random-scale local enhanced feature maps; and performing pseudo-texture avoidance on the filtered mapped local enhanced feature maps and filtered random-scale local enhanced feature maps to obtain mapped super-resolution local feature maps and random-scale super-resolution local feature maps.

[0016] Optionally, step S5 includes: aligning the fixed-scale candidate box with the mapped super-resolution local feature map and the random-scale super-resolution local feature map to obtain aligned fixed-scale candidate boxes with aligned mapped super-resolution local feature maps and random-scale super-resolution local feature maps; flattening the aligned fixed-scale candidate boxes with aligned mapped super-resolution local feature maps and random-scale super-resolution local feature maps to obtain fixed-scale feature vectors with mapped super-resolution local feature vectors and random-scale super-resolution local feature vectors; and applying a cosine similarity loss function to the fixed-scale feature vectors with mapped super-resolution local feature vectors and random-scale super-resolution local feature vectors. The semantic consistency loss value is calculated; the aligned mapped super-resolution local feature map and the aligned random scale super-resolution local feature map are input into the fine detection head of the detection model to predict the target candidate box, its coordinates, category, and category confidence; based on the predicted target candidate box, its coordinates, category, and category confidence, and the target calibration box, its coordinates, and category of the manually calibrated original UAV image, the category loss value and the second coordinate regression loss value are calculated; based on the semantic consistency loss value, the candidate box mapping loss value, the category loss value, and the second coordinate regression loss value, the target loss value is calculated; the parameters of the detection model are updated in reverse based on the target loss value.

[0017] On the other hand, this invention provides a progressive UAV detection system based on super-resolution guidance. The system includes: a feature extraction unit, used to acquire original UAV images, process them using both fixed and random resolutions, and input them into the backbone network of a detection model to extract fixed-scale and random-scale feature maps respectively; a feature enhancement unit, used to perform scale alignment and feature interaction between the random-scale feature map and the fixed-scale feature map to generate an enhanced fixed-scale feature map; a coarse prediction unit, used to predict fixed-scale candidate boxes and random-scale candidate boxes based on the enhanced fixed-scale feature map and the random-scale feature map respectively, and to map the fixed-scale candidate boxes to the random-scale feature map, calculating the candidate box mapping loss value based on the mapping result and the random-scale candidate boxes; and super-resolution enhancement... The strong unit is used to crop local feature maps from the fixed-scale feature map and the aligned random-scale feature map based on the mapped candidate boxes and random-scale candidate boxes, and perform super-resolution enhancement to obtain two super-resolution local feature maps; the reverse update unit is used to calculate the semantic consistency loss value of the fixed-scale candidate boxes and the two super-resolution local feature maps according to the cosine similarity loss function, and update the parameters of the detection model in combination with the candidate box mapping loss value; the repeated training unit is used to repeat the feature extraction unit, feature enhancement unit, coarse prediction unit, super-resolution enhancement unit, and reverse update unit until all loss values ​​converge to obtain the final detection model; the detection unit is used to acquire the image of the drone to be detected, process it only with a fixed resolution, and input it into the final detection model for detection, outputting the drone category and candidate box coordinates.

[0018] Optionally, the feature enhancement unit includes: an alignment subunit, used to perform scale alignment between the random-scale feature map and the fixed-scale feature map in the feature fusion module of the detection model to obtain an aligned feature map; a dot product subunit, used to perform element-wise dot product between the aligned feature map and the fixed-scale feature map to obtain a modulation feature map; a superposition subunit, used to perform element-wise superposition between the modulation feature map and the fixed-scale feature map to obtain a preliminary enhanced feature map; and an enhancement subunit, used to perform ReLU activation function operation on the preliminary enhanced feature map and perform batch normalization processing to obtain an enhanced fixed-scale feature map.

[0019] Optionally, the coarse prediction unit includes: a filtering subunit, used to predict target candidate boxes and their coordinates and confidence scores from the enhanced fixed-scale feature map and random-scale feature map respectively through the coarse detection head of the detection model; filtering the target candidate boxes according to a preset confidence threshold to remove background candidate boxes with confidence scores below the threshold, thereby obtaining fixed-scale candidate boxes and random-scale candidate boxes; a mapping subunit, used to calculate a scale conversion factor based on the ratio of fixed resolution to random resolution, and map the fixed-scale candidate boxes to the random-scale feature map according to the scale conversion factor to obtain mapped candidate boxes; a rejection subunit, used to perform boundary verification on the mapped candidate boxes and reject mapped candidate boxes that exceed the range of the random-scale feature map; a retention subunit, used to calculate the intersection-union ratio (IU) of the filtered mapped candidate boxes and the random-scale candidate boxes, and retain mapped candidate boxes with IU greater than or equal to a preset IU threshold; and a calculation subunit, used to calculate the candidate box mapping loss value based on the retained mapped candidate boxes and random-scale candidate boxes.

[0020] Optionally, the candidate box mapping loss value is calculated according to the following formula:

[0021] Lroi = α×L_reg+β×L_iou;

[0022] When |GT-Pred|≤1, L_reg=0.5×(GT-Pred)²; when |GT-Pred|>1, L_reg=|GT-Pred|-0.5;

[0023] L_iou = -ln(IOU);

[0024] Wherein, α is the first preset weight coefficient, β is the second preset weight coefficient, L_reg is the first coordinate regression loss value, L_iou is the intersection-union ratio loss value, Lroi is the candidate box mapping loss value, GT is the coordinates of the retained mapped candidate box, Pred is the coordinates of the random scale candidate box; and IOU is the mean of the intersection-union ratio of the retained mapped candidate box and the random scale candidate box.

[0025] The beneficial effects of this invention are:

[0026] This invention provides a progressive UAV detection method and system based on super-resolution guidance. The method employs a dual-resolution feature collaborative architecture, designing parallel branches for fixed and random scales. Through a feature fusion module involving scale alignment and feature interaction, it modulates and supplements fixed-scale features with random-scale features, directionally strengthening UAV discriminative features and effectively overcoming the deficiency of insufficient inter-module collaboration. A progressive process accurately locks the target region—first generating candidate boxes to focus on potential targets, then performing feature-level super-resolution enhancement only on the target region, fundamentally avoiding invalid processing of the background region and preventing the generation of false textures. This invention constructs a full-link semantic alignment system of "cross-scale mapping-loss constraint-pixel-level alignment": during the training phase, the cross-scale mapping accuracy is optimized through candidate box mapping loss (coordinate regression loss + IOU loss) to ensure spatial consistency of the target region under fixed and random scales; after super-resolution, alignment is performed, and a cosine similarity loss function is introduced to constrain the semantic consistency of the super-resolution features. This system effectively solves the problem of inaccurate localization caused by the lack of semantic alignment in traditional solutions. Through the synergistic effect of the aforementioned core mechanisms, this invention not only solves the fundamental problem of scarce features in low-resolution UAV images but also avoids the inherent defects of traditional solutions. Experiments show that the average accuracy of this method on public UAV detection datasets reaches 89.2%, which is 12.7 percentage points higher than the traditional "super-resolution + detection" concatenated framework. The inference stage retains only a fixed-scale branch, significantly reducing computational load and increasing inference speed by over 60%, thus meeting the routine real-time monitoring needs of ground-based monitors for aerial UAVs. Attached Figure Description

[0027] Figure 1 This is a flowchart of a progressive UAV detection method based on super-resolution guidance provided in an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of a progressive unmanned aerial vehicle (UAV) detection system based on super-resolution guidance provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0030] This embodiment uses a complete training and inference process as an example to illustrate the specific implementation of the technical solution of the present invention. For ease of understanding, the following specific parameters are set in this embodiment:

[0031] Raw drone image (i.e., raw low-resolution drone surveillance image) size: 100×100 pixels, 3-channel RGB format

[0032] Fixed resolution for fixed-scale branch input: 640×640

[0033] Random resolution of the random scale branch input: dynamically selected (randomly selected from a preset scale set {320×320, 512×512}), this embodiment takes 320×320 as an example.

[0034] Backbone network: ResNet50, downsampling factor 32x

[0035] Fixed-scale feature map size: 20×20×2048

[0036] Random scale feature map size: 10×10×2048

[0037] Training equipment: NVIDIA Tesla V100 GPU, batch size set to 32.

[0038] Optimizer: Stochastic Gradient Descent (SGD), initial learning rate 0.01, momentum 0.9, weight decay 0.0001

[0039] Training: The random resolution of the random scale branch is randomly switched every 10 epochs.

[0040] Figure 1 This is a flowchart of a progressive UAV detection method based on super-resolution guidance provided by an embodiment of the present invention; as shown below. Figure 1 As shown, the method includes:

[0041] S1. Acquire the original UAV image, process it using fixed resolution and random resolution respectively, and input it into the backbone network of the detection model to extract fixed-scale feature maps and random-scale feature maps respectively.

[0042] Specifically, during the model initialization phase, two parallel feature extraction branches are constructed, sharing the same input layer.

[0043] The input layer receives raw UAV images (i.e., raw low-resolution UAV images) from a ground monitor, with an image size of 100×100×3. For the fixed-scale branch, 640×640 is initially used as the fixed resolution. The input image is scaled to 640×640 using bilinear interpolation by the image preprocessing module, and then normalized with mean [0.485, 0.456, 0.406] and variance [0.229, 0.224, 0.225]. The normalized image is input into the ResNet50 backbone of the detection model (i.e., the UAV detection model). After progressive downsampling through five convolutional blocks (conv1-conv5, with downsampling ratios of 2, 2, 2, 2, 2 for a total downsampling ratio of 32), the final output is a fixed-scale feature map F_fix with a size of 20×20 (640÷32=20) and 2048 channels.

[0044] For the random scale branch, the scale scheduler selects a resolution of 320×320 (randomly selected from the preset scale set {320×320, 512×512}) in this training epoch (epochs 1-10). The image preprocessing module scales the original UAV image to 320×320 and processes it using the same normalization parameters as the fixed scale branch. The random scale branch also uses ResNet50 as the backbone network and shares network weights with the fixed scale branch. After the input image is input at a resolution of 320×320, it undergoes the same downsampling process, outputting a random scale feature map F_rand with a size of 10×10×2048.

[0045] The forward propagation of the two branches is executed in parallel, and the output feature maps are stored in the feature cache unit for subsequent modules to call.

[0046] S2 performs scale alignment and feature interaction between the random scale feature map and the fixed scale feature map to generate an enhanced fixed scale feature map.

[0047] In an optional implementation, S2 includes:

[0048] S21. In the feature fusion module of the detection model, the random scale feature map and the fixed scale feature map are scale-aligned to obtain an aligned feature map;

[0049] Specifically, the feature fusion module receives a fixed-scale feature map F_fix (20×20×2048) and a random-scale feature map F_rand (10×10×2048) as input.

[0050] First, scale alignment is performed: the scaling factor k = 20 / 10 = 2 is calculated, and the random scale feature map F_rand is upsampled by a factor of 2 using bilinear interpolation to obtain the scaled feature map (i.e., the aligned feature map) F_rand_align, with a size of 20×20×2048. During the upsampling process, edge padding (with the padding mode set to "reflection padding") is used to prevent the loss of edge features during scaling.

[0051] S22. Perform element-wise multiplication between the aligned feature map and the fixed-scale feature map to obtain the modulation feature map;

[0052] Specifically, feature interaction is performed: an element-wise multiplication operation is performed between the aligned feature map F_rand_align and the fixed-scale feature map F_fix. This achieves weight modulation of the fixed-scale feature by the random-scale feature—positions with larger values ​​in the multiplication result correspond to common effective regions of both types of features (such as UAV contour features), strengthening the target feature response; positions with smaller values ​​correspond to background or noise regions, weakening invalid information; for each spatial position (i,j) and each channel c, the modulated feature map F_mul(i,j,c) = F_rand_align(i,j,c) × F_fix(i,j,c) is calculated. This operation strengthens regions where both branches respond jointly (i.e., regions where both branches believe a target may exist), while regions with strong responses from a single branch (which may be noise or background) are suppressed.

[0053] S23. The modulation feature map is superimposed element-wise with the fixed-scale feature map to obtain a preliminary enhanced feature map;

[0054] Next, residual fusion is performed: the modulated feature map F_mul is element-wise superimposed with the fixed-scale feature map F_fix. Before superposition, a 1×1 convolutional layer is used to adapt the channels of the fixed-scale feature map F_fix. Although both branches have 2048 channels, this convolutional layer provides additional non-linear transformation capabilities, improving the model's generalization ability. After superposition, the preliminary enhanced feature map F_add = Conv1×1(F_fix) + F_mul is obtained.

[0055] S24. Perform ReLU activation function operation on the preliminary enhanced feature map and obtain the enhanced fixed-scale feature map through batch normalization.

[0056] Finally, the initial enhanced feature map F_add is subjected to ReLU activation function operation (introducing non-linear expressive power and suppressing gradient vanishing), and then batch normalization is performed. The momentum parameter of batch normalization is set to 0.9 and epsilon is set to 1e-5 to stabilize the feature distribution. The processed feature map is the enhanced fixed-scale feature map F_enh, with the size maintained at 20×20×2048.

[0057] Thus, the enhanced fixed-scale feature map F_enh retains the strong semantic information of the original fixed-scale branch (the advantage of low-resolution features) and incorporates the detailed features provided by the random-scale branch (the advantage of high-resolution features), achieving deep complementary enhancement of high and low resolution features.

[0058] S3. Predict fixed-scale candidate boxes and random-scale candidate boxes based on the enhanced fixed-scale feature map and random-scale feature map respectively, and map the fixed-scale candidate boxes to the random-scale feature map. Calculate the candidate box mapping loss value based on the mapping result and the random-scale candidate boxes. In this step, coarse target detection is performed on the enhanced fixed-scale feature map F_enh (size 20×20×2048) and the random-scale feature map F_rand (size 10×10×2048) respectively to generate candidate boxes. Spatial alignment of the dual-branch target regions is achieved through coordinate mapping and loss calculation, providing accurate localization for subsequent local super-resolution enhancement.

[0059] In an optional implementation, S3 includes:

[0060] S31. The enhanced fixed-scale feature map and random-scale feature map are respectively used by the coarse detection head of the detection model to predict the target candidate box and its coordinates and confidence. The target candidate box is filtered according to the preset confidence threshold to remove the background candidate box with a confidence lower than the threshold, so as to obtain the fixed-scale candidate box and the random-scale candidate box.

[0061] Two identical coarse detection heads are constructed and connected to the enhanced fixed-scale feature map F_enh and the random-scale feature map F_rand, respectively. Each detection head employs a predefined anchor mechanism: three anchor boxes of different scales and three different aspect ratios are pre-defined at each feature map location (a total of nine anchor boxes). Two parallel 1×1 convolutional layers predict the confidence score of each anchor box containing the target and its positional offset relative to the anchor box. After decoding, a set of candidate boxes is obtained, each containing position coordinates (x1, y1, x2, y2) (feature map coordinate system), target confidence score, and class probability. A confidence threshold θ_conf = 0.5 is set to remove background candidate boxes with confidence scores below the threshold.

[0062] Fixed-scale branch: Multiple candidate boxes are predicted on the enhanced fixed-scale feature map F_enh. After confidence filtering, two high-confidence candidate boxes are retained (denoted as B1 and B2):

[0063] B1: Coordinates (10, 10, 18, 18), confidence level 0.92

[0064] B2: Coordinates (5, 12, 10, 18), confidence level 0.78

[0065] Random scaling branch: After predicting and filtering on F_rand, retain 3 candidate boxes (denoted as C1, C2, C3):

[0066] C1: Coordinates (4.2, 4.1, 8.3, 7.8), confidence level 0.88 (the coordinates here are continuous values, directly output by the detection head, and are not rounded)

[0067] C2: Coordinates (6.3, 5.2, 9.1, 8.9), confidence level 0.76

[0068] C3: Coordinates (2.1, 3.2, 5.3, 6.1), confidence level 0.81

[0069] S32. Calculate the scale conversion factor based on the ratio of fixed resolution to random resolution, and map the fixed-scale candidate box to the random-scale feature map based on the scale conversion factor to obtain the mapped candidate box.

[0070] Since the fixed-scale feature map has a fixed resolution of 20×20, and the random-scale feature map has a random resolution of 10×10, a scale conversion factor is defined as follows:

[0071] Width conversion factor k_w = 10 / 20 = 0.5, Height conversion factor k_h = 10 / 20 = 0.5

[0072] Multiplying the coordinates of the fixed-scale candidate boxes by this factor yields the theoretical mapped coordinates (continuous values):

[0073] Theoretical coordinates after B1 mapping: (5.0, 5.0, 9.0, 9.0)

[0074] Theoretical coordinates after B2 mapping: (2.5, 6.0, 5.0, 9.0)

[0075] To facilitate subsequent feature clipping, nearest neighbor interpolation (rounded) is used to obtain integer coordinates: M1 (mapped from B1): (5,5,10,10)

[0076] M2 (mapped from B2): (3,6,5,9) (2.5 is rounded to 3, 6.0 is rounded to 6, 5.0 is rounded to 5, and 9.0 is rounded to 9)

[0077] That is, M1 and M2 are the mapping candidate boxes.

[0078] S33. Perform boundary verification on the mapping candidate boxes and remove mapping candidate boxes that exceed the range of the random scale feature map;

[0079] Check if the integer coordinates are within the valid range [0,9] of the random scale feature map. Both M1 and M2 coordinates are within the range, so both mapping candidate boxes pass the verification.

[0080] S34. Calculate the intersection-union ratio (IU) of the filtered mapped candidate boxes and the random scale candidate boxes, and retain the mapped candidate boxes whose IU is greater than or equal to the preset IU threshold.

[0081] To evaluate the spatial consistency between mapped candidate boxes and random-scale candidate boxes, it is necessary to restore both to the original image coordinate system (original image size 100×100) and calculate the Interchange of Units (IOU). The transformation relationship is as follows:

[0082] Each pixel in the fixed-scale feature map corresponds to 5 pixels in the original drone image (100 / 20=5).

[0083] Each pixel in the random scale feature map corresponds to 10 pixels in the original drone image (100 / 10=10).

[0084] Transform each candidate bounding box to the original drone image coordinates:

[0085] M1 (integer coordinates (5,5,9,9)): (50,50,90,90)

[0086] M2 (integer coordinates (3,6,5,9)): (30,60,50,90)

[0087] C1 (continuous coordinates (4.2,4.1,8.3,7.8)): First round to (4,4,8,8), then convert to (40,40,80,80) (For simplification, the calculation is performed after rounding; continuous values ​​are still used when calculating the actual loss).

[0088] C2: Rounding down to (6,5,9,9) → (60,50,90,90)

[0089] C3: Rounding down to (2,3,5,6) → (20,30,50,60)

[0090] Calculate the Intersection over Union (IOU) of all mapped candidate boxes with all random scale candidate boxes in the original image space to obtain 6 Intersection over Union ratios (IoU). Retain mapped candidate boxes with IoU ≥ 0.7 to ensure the consistency of target regions across scales.

[0091] Assuming that all six intersection-union ratios are greater than 0.7, we obtain the retained mapping candidate boxes M1, M2 and random scale candidate boxes C1, C2, C3.

[0092] S35. Calculate the candidate box mapping loss value based on the retained mapping candidate boxes and random scale candidate boxes.

[0093] In an optional implementation, the candidate box mapping loss value is calculated according to the following formula:

[0094] Lroi = α×L_reg+β×L_iou;

[0095] When |GT-Pred|≤1, L_reg=0.5×(GT-Pred)²; when |GT-Pred|>1, L_reg=|GT-Pred|-0.5;

[0096] L_iou = -ln(IOU);

[0097] Wherein, α is the first preset weight coefficient, β is the second preset weight coefficient, L_reg is the first coordinate regression loss value, L_iou is the intersection-union ratio loss value, Lroi is the candidate box mapping loss value, GT is the coordinates of the retained mapped candidate box, Pred is the coordinates of the random scale candidate box; and IOU is the mean of the intersection-union ratio of the retained mapped candidate box and the random scale candidate box.

[0098] Specifically, using candidate boxes C1, C2, and C3 at random scale as a benchmark, determine which has the largest intersection-union ratio (IU) between C1 and M1, and between C1 and M2. If the IU of C1 and M1 is larger, Then calculate the coordinate regression loss value of C1 and M1, where the coordinates of M1 are GT and the coordinates of C1 are Pred. Similarly, determine which is larger, the intersection-union ratio of C2 and M1 or the intersection-union ratio of C2 and M2. If the intersection-union ratio of C2 and M2 is larger, then calculate the coordinate regression loss value of C2 and M2, where the coordinates of M2 are GT and the coordinates of C2 are Pred. Determine which is larger, the intersection-union ratio of C3 and M1 or the intersection-union ratio of C3 and M2. If the intersection-union ratio of C3 and M1 is larger, then calculate the coordinate regression loss value of C3 and M1, where the coordinates of M1 are GT and the coordinates of C3 are Pred. Calculate the average of the three coordinate regression losses to obtain the first coordinate regression loss value.

[0099] IOU is the mean of the sums of the intersection-union ratios of C1 and M1, C1 and M2, C2 and M1, C2 and M2, C3 and M1, and C3 and M2.

[0100] S4. Based on the mapped candidate box and the random scale candidate box, crop the local feature map from the fixed scale feature map and the aligned random scale feature map, and perform super-resolution enhancement on it to obtain two super-resolution local feature maps.

[0101] This step, based on the candidate bounding boxes M1, M2 and random scale candidate bounding boxes C1, C2, C3 selected and retained in step three, cropes the corresponding local regions from the fixed scale feature map F_enh and the random scale feature map F_rand, respectively. Then, the local super-resolution module performs feature-level enhancement on each local region to strengthen the discriminative features of the UAV target while suppressing noise and pseudo-textures.

[0102] In an optional implementation, S4 includes:

[0103] S41. Based on the coordinates of the retained mapping candidate box and the random scale candidate box, crop out the corresponding local feature maps from the fixed scale feature map and the random scale feature map after aligning with the fixed scale feature map.

[0104] To facilitate subsequent cropping operations, the random-scale feature map F_rand (size 10×10×2048) is first upsampled to match the spatial dimensions of the fixed-scale feature map F_enh (size 20×20×2048). A bilinear interpolation algorithm with a scaling factor k=2 is then used to obtain the aligned random-scale feature map F_rand_align, with a size of 20×20×2048. Reflection padding is employed during the upsampling process to prevent the loss of edge features.

[0105] Since the coordinates of the mapped candidate boxes M1 and M2 in step three are obtained through nearest neighbor interpolation and rounding, there may be pixel-level offsets. Therefore, fine-tuning and calibration are required. The calibration formula is:

[0106] x_cal=round(x_pred+0.5), y_cal=round(y_pred+0.5)

[0107] Where x_pred and y_pred are the theoretical coordinates of the mapped candidate boxes, and round is the rounding function.

[0108] For the mapping candidate box M1, its theoretical coordinates are (5.0, 5.0, 9.0, 9.0), and after calibration, they remain (5, 5, 9, 9); for M2, the theoretical coordinates are (2.5, 6.0, 5.0, 9.0), and after calibration, they are (3, 6, 5, 9). Transforming the calibrated coordinates to a 20×20 coordinate system (multiplied by 2) yields the clipping coordinates on F_enh:

[0109] M1: (10,10,18,18)

[0110] M2: (6,12,10,18)

[0111] For candidate bounding boxes C1, C2, and C3 of random scale, their coordinates are directly output as continuous values ​​by the detector head, and are also fine-tuned, calibrated, and transformed to a 20×20 coordinate system:

[0112] The continuous coordinates in C1 are (4.2, 4.1, 8.3, 7.8), which, after calibration, are (5, 5, 9, 8). Mapping this to 20×20, we get (10, 10, 18, 16).

[0113] The C2 continuous coordinates (6.3, 5.2, 9.1, 8.9), after calibration (7, 6, 10, 9), are mapped to (14, 12, 20, 18).

[0114] The C3 continuous coordinates (2.1, 3.2, 5.3, 6.1), after calibration (3, 4, 6, 7), are mapped to (6, 8, 12, 14).

[0115] Based on the calibrated coordinates, the corresponding local feature maps are cropped from F_enh and F_rand_align respectively:

[0116] By cropping region M1 from F_enh, we obtain local feature map P1 with size (18). 10) = 8 rows, (18) 10) = 8 columns, that is, 8×8×2048.

[0117] By cropping region M2 from F_enh, we obtain local feature map P2 with size (10). 6) = 4 rows, (18) 12) = 6 columns, so the dimensions are 4×6×2048.

[0118] Cropping region C1 from F_rand_align yields local feature map Q1 with size (18). 10) = 8 rows, (16) 10) = 6 columns, that is, 8 × 6 × 2048.

[0119] Cropping region C2 from F_rand_align yields Q2, with a size of (20). 14) = 6 rows, (18) 12) = 6 columns, that is, 6×6×2048.

[0120] Cropping region C3 from F_rand_align yields Q3, with a size of (12... 6) = 6 rows, (14) 8) = 6 columns, that is, 6×6×2048.

[0121] During the cropping process, if the target region exceeds the feature map boundary (e.g., coordinates less than 0 or greater than 19), edge completion is performed. During completion, for pixels exceeding the boundary, bilinear interpolation weighting is applied to the effective feature values ​​within their 3×3 neighborhood. The weighting coefficient decreases linearly with distance to ensure the continuity of edge features. In this example, all cropped regions are within the effective range, so no completion is needed.

[0122] S42. Input the cropped local feature map into the local super-resolution module for feature-level super-resolution enhancement to obtain the mapped local enhanced feature map and the random-scale local enhanced feature map; the local super-resolution module includes a channel attention sub-module and a residual learning sub-module;

[0123] Each cropped local feature map is input into the local super-resolution module for feature-level enhancement. The local super-resolution module adopts a composite structure of "SE channel attention sub-module + 3-layer residual block", which can adaptively enhance key feature channels and supplement detailed information. The enhancement process is described in detail below using P1 (8×8×2048) as an example.

[0124] (1) SE channel attention submodule

[0125] This submodule aims to learn the importance weights of each feature channel and strengthen the important channels to highlight discriminative features such as drone outline and texture.

[0126] Global average pooling: Global average pooling is performed on the input feature map P1, compressing the spatial dimension from 8×8 to 1×1, resulting in the channel statistics vector z∈R. 2048 ;

[0127] Fully connected layer learning weights: z is input into a two-layer fully connected network. The first fully connected layer (FC1) compresses the 2048 dimensions to 512 dimensions (compression ratio r=4), with ReLU activation function; the second fully connected layer (FC2) restores the 2048 dimensions, with Sigmoid activation function, and outputs a channel weight vector w∈[0,1]. 2048 This indicates the importance of each channel.

[0128] Channel-weighted: The weight vector w is multiplied channel by channel with the original feature map to obtain the attention-enhanced feature map P. att :

[0129] P att (i,j,c)=w·P1(i,j,c)

[0130] This operation enhances channels that contribute significantly to the drone's discriminative features (such as edges and textures) while suppressing background or noise channels.

[0131] After channel attention enhancement, feature map P att An input residual learning submodule is used to further supplement high-frequency details. This submodule consists of three residual blocks cascaded together. Each residual block contains two 3×3 convolutional layers (2048 kernels, stride 1, padding=1), followed by batch normalization (BN, momentum 0.9, epsilon=1e-5) and ReLU activation after each convolution. The output of the l-th residual block can be represented as:

[0132] F l =ReLU(BN(Conv 3×3 (F l 1)))⊕F l 1

[0133] Where ⊕ represents element-wise addition (residual join). Input F0=P att P is obtained after passing through 3 layers of residual blocks. res The dimensions remain 8×8×2048.

[0134] Residual connections ensure the preservation of original semantic information, while convolutional layers supplement detailed features, making the enhanced features more discriminative. For other local feature maps (P2, Q1, Q2, Q3), the exact same submodule structure is used for processing, with the output size being the same as the input. Because the convolutional layers use padding=1, convolution operations can still be performed normally for non-square or small-sized feature maps, and the output size remains unchanged.

[0135] S43. Noise filtering is performed on the mapped local augmentation feature map and the random scale local augmentation feature map;

[0136] The enhanced feature map (i.e., the mapped local enhanced feature map and the random-scale local enhanced feature map) may contain noise (such as artifacts caused by the super-resolution module amplification or noise in the original features), which needs to be further filtered. This step uses a method that combines adaptive noise detection and neighborhood smoothing.

[0137] (1) Noise detection

[0138] For each enhanced feature map (taking P res as an example), calculate the average gradient magnitude at each spatial position (i, j) across all channels. First, use the Sobel operator to calculate the horizontal gradient Gx and vertical gradient Gy (convolution kernel size 3×3) to obtain the gradient map for each channel; then calculate the gradient magnitude:

[0139] ;

[0140] Set the threshold T = 0.05 (empirical value, which can be fine-tuned according to the noise level of the training data). If G(i, j) < T, then this position is considered a noise pixel.

[0141] (2) Noise filtering

[0142] Adaptive threshold filtering is applied to the noise pixels to obtain the filtered mapped local enhanced feature map and the filtered random-scale local enhanced feature map. Taking Pres as an example, the corresponding filtered mapped local enhanced feature map is P filter .

[0143] (3) Neighborhood smoothing

[0144] Perform 3×3 Gaussian smoothing on the filtered feature map P filter , with the standard deviation σ of the Gaussian kernel being 0.8, to further suppress Gaussian noise and salt-and-pepper noise and ensure the smoothness of the feature map and the integrity of the target features. After smoothing, the feature map F1 is obtained.

[0145] S44. Perform pseudo-texture avoidance on the filtered mapped local enhanced feature map and the filtered random-scale local enhanced feature map to obtain the mapped super-resolved local feature map and the random-scale super-resolved local feature map.

[0146] To prevent the super-resolution module from generating false textures that do not conform to the true UAV features, a UAV feature prior library is introduced for constraint.

[0147] (1) Prior constraint

[0148] The UAV feature prior library contains typical feature vectors extracted from a large number of real UAV images (obtained by extraction and clustering using the same backbone network), each vector having a dimension of 2048 and covering UAV contours and texture features under different poses, scales, and lighting conditions.

[0149] Perform global average pooling on the smoothed feature map F1 to obtain the global feature vector v ∈ R 2048 . Calculate the cosine similarity between v and all template vectors in the prior library, and take the maximum value sim max :

[0150] Set the similarity threshold θ prior =0.6. If sim max <θ prior This indicates that the current features deviate from the real drone features and there may be pseudo textures. In this case, the super-resolution module parameters need to be adjusted and regenerated (for example, slightly adjust the convolution kernel weights or attention weights in the residual block and re-execute S42 and S43).

[0151] If the similarity meets the requirements, the current feature map F1 is used as the enhanced local feature map of the region, that is, the local feature map after super-resolution mapping.

[0152] (2) Standardization

[0153] Since the local feature maps cropped from different candidate boxes have different sizes (e.g., 8×8, 4×6, 6×6, etc.), to facilitate semantic consistency calculation in step five, all enhanced local feature maps need to be uniformly adjusted to a fixed size of 10×10×2048. Bilinear interpolation is used for upsampling or downsampling.

[0154] For P1 (8×8), upsample to 10×10;

[0155] For P2 (4×6), first upsample to 10×10 (independent interpolation in each direction);

[0156] For Q1 (8×6), Q2 (6×6), and Q3 (6×6), interpolate to 10×10.

[0157] Bilinear interpolation is used during the interpolation process to ensure a smooth transition of pixel values.

[0158] (3) Feature output

[0159] After scaling, we obtain the final local feature maps after super-resolution and the local feature maps after random-scale super-resolution:

[0160] Local feature maps after super-resolution mapping: F1 (derived from P1) and F2 (derived from P2), both with a size of 10×10×2048.

[0161] Local feature maps after random scale super-resolution: F3 (derived from Q1), F4 (derived from Q2), F5 (derived from Q3), all with a size of 10×10×2048.

[0162] These enhanced local feature maps (i.e., the local feature maps after mapping super-resolution and the local feature maps after random scale super-resolution) contain only real UAV target features and are free from background pseudo-texture interference, providing high-quality feature input for the semantic consistency loss calculation and final target detection in the subsequent step five.

[0163] At this point, step four is complete, and feature-level super-resolution enhancement has been completed for the local regions corresponding to all candidate boxes.

[0164] S5. Calculate the semantic consistency loss value between the fixed-scale candidate boxes and the two super-resolution local feature maps using the cosine similarity loss function, and update the parameters of the detection model by combining the candidate box mapping loss value.

[0165] In an optional implementation, S5 includes:

[0166] S51. Align the fixed-scale candidate box with the local feature map after mapping and the local feature map after random-scale super-resolution to obtain the aligned fixed-scale candidate box with the aligned local feature map after mapping and the local feature map after random-scale super-resolution.

[0167] Specifically, the corresponding local regions of the fixed-scale candidate boxes B1 and B2 are cropped from the fixed-scale feature map F_enh to obtain the corresponding local feature maps, namely D1 and D2.

[0168] Pixel-level alignment of D1, D2 with F1, F2, F3, F4, and F5 is performed using bilinear interpolation, resulting in aligned fixed-scale candidate boxes and aligned mapped super-resolution local feature maps, as well as random-scale super-resolution local feature maps.

[0169] S52. Flatten the aligned fixed-scale candidate boxes with the aligned mapped super-resolution local feature maps and random-scale super-resolution local feature maps to obtain fixed-scale feature vectors, mapped super-resolution local feature vectors, and random-scale super-resolution local feature vectors.

[0170] S53. The semantic consistency loss value is calculated by combining the fixed-scale feature vector with the local feature vector after mapping and the local feature vector after random-scale super-resolution using the cosine similarity loss function.

[0171] Specifically, after alignment and conversion into feature vectors, D1 is compared with F1, F2, F3, F4, and F5 to obtain 5 similarity values. After alignment and conversion into feature vectors, D2 is compared with F1, F2, F3, F4, and F5 to obtain 5 similarity values.

[0172] Sum the 10 similarity values ​​and take the mean to get the mean similarity value. 1 - mean similarity value = semantic consistency loss value.

[0173] S54. Input the aligned mapped super-resolution local feature map and the aligned random scale super-resolution local feature map into the fine detection head of the detection model to predict the target candidate box, its coordinates, category and category confidence; calculate the category loss value and the second coordinate regression loss value based on the predicted target candidate box, its coordinates, category and category confidence, as well as the target calibration box, its coordinates and category of the manually calibrated original UAV image.

[0174] Specifically, a category loss value is calculated based on the predicted target candidate boxes, their categories and category confidence, as well as the manually calibrated target bounding boxes and their categories in the original UAV image; a second coordinate loss value is calculated based on the predicted target candidate boxes, their coordinates, and the manually calibrated target bounding boxes and their coordinates in the original UAV image.

[0175] S55. Calculate the target loss value based on the semantic consistency loss value, candidate box mapping loss value, category loss value, and second coordinate regression loss value; update the parameters of the detection model in reverse based on the target loss value.

[0176] Specifically, the semantic consistency loss value, candidate box mapping loss value, category loss value, and second coordinate regression loss value are added together to obtain the target loss value; the parameters of the detection model are then updated in reverse based on the target loss value.

[0177] S6. Repeat S1~S5 until all loss values ​​and values ​​converge to obtain the final detection model;

[0178] Repeat steps S1 to S5, i.e., train multiple times, until the target loss value is within the first preset range ( The training stops when the model fluctuates within a certain range, resulting in the final detection model.

[0179] S7. Obtain the image of the drone to be detected, process it using only a fixed resolution, and input it into the final detection model for detection, outputting the drone category and candidate box coordinates.

[0180] Specifically, the model parameters are loaded from the final detection model saved during the training phase, including:

[0181] Convolutional layer parameters (conv1-conv5) of the backbone network ResNet50.

[0182] The weights of the 1×1 convolutional layer in the feature fusion module, and the mean and variance of the batch normalization layer;

[0183] The channel attention submodule (fully connected layer weights) and residual blocks (convolutional kernel weights and BN layer parameters of 3 residual blocks) of the local super-resolution module.

[0184] Parameters of the classification and regression layers of the detection head (including coarse and fine detection heads, with only the fine detection head used during inference);

[0185] After loading is complete, initialize the fixed-scale feature branch and set the input fixed resolution to 640×640.

[0186] Suppose a ground monitor acquires a new low-resolution image of a drone in real time, with an original size of 100×100×3 (3-channel RGB format), denoted as I. orig .

[0187] First, the bilinear interpolation algorithm is used to... orig Scaled to a fixed size of 640×640, we get image I. resize Next, regarding I... resize Standardization was performed using the same normalization parameters as in the training phase: mean: μ=[0.485,0.456,0.406]; variance: σ=[0.229,0.224,0.225].

[0188] Standardized image I norm The pixel value distribution is consistent with the training data, eliminating the interference of magnitude difference on model inference, and can be directly input into the backbone network.

[0189] Standardized image I norm The input is a ResNet50 backbone network with a fixed-scale branch. The backbone network contains 5 convolutional blocks (conv1-conv5), each with a downsampling factor of 2, 2, 2, 2, and 2 respectively, for a total downsampling factor of 32. The final output is a 20×20×2048 fixed-scale feature map to be detected. Since there is no random-scale branch in the inference stage, the optimized feature fusion module parameters from the training stage are directly called to perform residual enhancement and normalization on the fixed-scale feature map to be detected, outputting an enhanced feature map to be detected that retains strong semantic information and detailed features, ensuring the discriminability of the target features.

[0190] The enhanced feature map to be detected is used to predict target candidate boxes (including location coordinates (x1, y1, x2, y2), category and category confidence) through the coarse detection head of the final detection model; Confidence screening: Set the confidence threshold (0.5) consistent with the training stage, and remove background candidate boxes with confidence scores lower than the confidence threshold to initially screen potential drone targets;

[0191] The retained target candidate boxes are extracted from the fixed-scale feature map; the cropped local feature map is input into the local super-resolution module for feature-level super-resolution enhancement to obtain the fixed-scale local enhanced feature map to be detected; noise is filtered from the fixed-scale local enhanced feature map to be detected, and pseudo-texture avoidance is performed on the filtered fixed-scale local enhanced feature map to be detected to obtain the fixed-scale super-resolution local feature map to be detected.

[0192] The local feature map after fixed-scale super-resolution is used to detect the drone category and candidate box coordinates by passing it through the fine detection head of the final detection model.

[0193] Through the above steps, this invention realizes a complete process from training to inference. During the training phase, it learns powerful detection capabilities through mechanisms such as dual-branch collaboration, local super-resolution, and semantic alignment. During the inference phase, it completes efficient and accurate UAV detection with minimal fixed-scale branches, thus solving the contradiction between accuracy and speed in UAV detection in low-resolution images.

[0194] Figure 2 This is a schematic diagram of a progressive unmanned aerial vehicle (UAV) detection system based on super-resolution guidance provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the system includes:

[0195] The feature extraction unit 201 is used to acquire the original UAV image, process it with fixed resolution and random resolution respectively, and input it into the backbone network of the detection model to extract fixed-scale feature maps and random-scale feature maps respectively.

[0196] The feature enhancement unit 202 is used to perform scale alignment and feature interaction between the random scale feature map and the fixed scale feature map to generate an enhanced fixed scale feature map.

[0197] In an optional implementation, the feature enhancement unit 202 includes:

[0198] Alignment subunit 2021 is used to perform scale alignment between the random scale feature map and the fixed scale feature map in the feature fusion module of the detection model to obtain an aligned feature map;

[0199] The dot product subunit 2022 is used to perform element-wise dot product between the aligned feature map and the fixed-scale feature map to obtain the modulation feature map;

[0200] The superposition subunit 2023 is used to superimpose the modulation feature map and the fixed-scale feature map element by element to obtain a preliminary enhanced feature map;

[0201] The enhancement subunit 2024 is used to perform ReLU activation function operation on the preliminary enhanced feature map and obtain the enhanced fixed-scale feature map through batch normalization.

[0202] The coarse prediction unit 203 is used to predict fixed-scale candidate boxes and random-scale candidate boxes based on the enhanced fixed-scale feature map and random-scale feature map, respectively, and to map the fixed-scale candidate boxes to the random-scale feature map. The candidate box mapping loss value is calculated based on the mapping result and the random-scale candidate boxes.

[0203] In an optional implementation, the coarse prediction unit 203 includes:

[0204] The filtering subunit 2031 is used to predict target candidate boxes and their coordinates and confidence scores by passing the enhanced fixed-scale feature map and random-scale feature map through the coarse detection head of the detection model; and to filter the target candidate boxes according to a preset confidence threshold to remove background candidate boxes with confidence scores lower than the threshold, thereby obtaining fixed-scale candidate boxes and random-scale candidate boxes.

[0205] The mapping subunit 2032 is used to calculate the scale conversion factor based on the ratio of fixed resolution and random resolution, and to map the fixed-scale candidate box to the random-scale feature map based on the scale conversion factor to obtain the mapped candidate box.

[0206] The elimination subunit 2033 is used to perform boundary verification on the mapping candidate boxes and eliminate mapping candidate boxes that exceed the range of the random scale feature map.

[0207] The retention subunit 2034 is used to calculate the intersection-union ratio (IU) between the filtered mapping candidate boxes and the random-scale candidate boxes, and to retain mapping candidate boxes whose IU is greater than or equal to a preset IU threshold.

[0208] Computation subunit 2035 is used to calculate the candidate box mapping loss value based on the retained mapping candidate boxes and random scale candidate boxes.

[0209] In an optional implementation, the candidate box mapping loss value is calculated according to the following formula:

[0210] Lroi = α×L_reg+β×L_iou;

[0211] When |GT-Pred|≤1, L_reg=0.5×(GT-Pred)²; when |GT-Pred|>1, L_reg=|GT-Pred|-0.5;

[0212] L_iou = -ln(IOU);

[0213] Wherein, α is the first preset weight coefficient, β is the second preset weight coefficient, L_reg is the first coordinate regression loss value, L_iou is the intersection-union ratio loss value, Lroi is the candidate box mapping loss value, GT is the coordinates of the retained mapped candidate box, Pred is the coordinates of the random scale candidate box; and IOU is the mean of the intersection-union ratio of the retained mapped candidate box and the random scale candidate box.

[0214] The super-resolution enhancement unit 204 is used to crop local feature maps from fixed-scale feature maps and aligned random-scale feature maps according to the mapped candidate boxes and random-scale candidate boxes, and perform super-resolution enhancement on them to obtain two super-resolution local feature maps.

[0215] The reverse update unit 205 is used to calculate the semantic consistency loss value of the fixed-scale candidate box and the two super-resolution local feature maps according to the cosine similarity loss function, and update the parameters of the detection model by combining the candidate box mapping loss value.

[0216] The repeated training unit 206 is used to repeat the feature extraction unit, feature enhancement unit, coarse prediction unit, super-resolution enhancement unit, and reverse update unit until all loss values ​​and values ​​converge to obtain the final detection model.

[0217] The detection unit 207 is used to acquire images of the drone to be detected, process them using only a fixed resolution, and input them into the final detection model for detection, outputting the drone category and candidate box coordinates.

[0218] The system of the present invention corresponds to the method described above, and the specific implementation of the system will not be repeated here.

[0219] Compared with the prior art, the present invention has the following beneficial effects:

[0220] 1. Break through the barriers of "independent modules" and build a deep collaborative mechanism.

[0221] In existing technologies, super-resolution reconstruction and object detection are used as independent modules in series, with no information interaction between them, and the features required for detection are not targeted for enhancement. This invention, through a dual-resolution feature collaborative architecture, designs parallel branches for fixed and random scales. A feature fusion module, employing "scale alignment - feature interaction - residual fusion," modulates and supplements fixed-scale features with random-scale features, targeting and enhancing the UAV's discriminative features, effectively overcoming the shortcomings of insufficient inter-module collaboration. During the training phase, weight sharing reduces parameter redundancy, and feature-guided learning optimizes collaborative accuracy. During the inference phase, the fusion module parameters are reused, balancing detection performance and real-time performance.

[0222] 2. Innovative local super-resolution strategy to avoid pseudo-texture interference.

[0223] Existing technologies perform super-resolution reconstruction on the entire image, which is not only computationally intensive but also prone to generating false textures that lead to misjudgments. This invention proposes a "feature-level local super-resolution" concept, which precisely locates the target region through a progressive process—first generating candidate boxes to focus on potential targets, and then performing feature-level super-resolution enhancement only on the target region. This fundamentally avoids ineffective processing of the background region and prevents the generation of false textures. The local super-resolution module integrates channel attention mechanisms and residual learning, effectively filtering out interference through a noise suppression unit while enhancing target features, making the enhanced features more closely resemble the characteristics of a real drone.

[0224] 3. Establish a semantic alignment system to overcome positioning bias.

[0225] Existing technologies lack cross-module spatial alignment mechanisms, leading to target displacement after super-resolution and resulting in detection and localization errors. This invention constructs a full-link semantic alignment system of "cross-scale mapping - loss constraint - pixel-level alignment": during the training phase, the cross-scale mapping accuracy is optimized through candidate box mapping loss (coordinate regression loss + IOU loss) to ensure spatial consistency of the target region under fixed and random scales; after super-resolution, pixel-level fine-tuning achieves precise alignment, and cosine similarity loss is introduced to constrain the semantic consistency of features after super-resolution. This system effectively solves the problem of inaccurate localization caused by the lack of semantic alignment in traditional solutions.

[0226] 4. Significantly improves detection accuracy and efficiency

[0227] Through the synergistic effect of the aforementioned core mechanisms, this invention not only solves the fundamental problem of scarce features in low-resolution UAV images but also avoids the inherent defects of traditional solutions. Experiments show that the average accuracy of this method on public UAV detection datasets reaches 89.2%, which is 12.7 percentage points higher than the traditional "super-resolution + detection" concatenated framework. The inference stage retains only a fixed-scale branch, significantly reducing computational load and increasing inference speed by over 60%, thus meeting the routine real-time monitoring needs of ground-based monitors for aerial UAVs.

[0228] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A progressive UAV detection method based on super-resolution guidance, characterized in that, include: S1. Acquire the original UAV image, process it using fixed resolution and random resolution respectively, and input it into the backbone network of the detection model to extract fixed-scale feature maps and random-scale feature maps respectively. S2. Align the random scale feature map with the fixed scale feature map and perform feature interaction to generate an enhanced fixed scale feature map. S3. Predict fixed-scale candidate boxes and random-scale candidate boxes based on the enhanced fixed-scale feature map and random-scale feature map respectively, and map the fixed-scale candidate boxes to the random-scale feature map. Calculate the candidate box mapping loss value based on the mapping result and the random-scale candidate boxes. S4. Based on the mapped candidate box and the random scale candidate box, crop the local feature map from the fixed scale feature map and the aligned random scale feature map, and perform super-resolution enhancement on it to obtain two super-resolution local feature maps. S5. Calculate the semantic consistency loss value between the fixed-scale candidate boxes and the two super-resolution local feature maps using the cosine similarity loss function, and update the parameters of the detection model by combining the candidate box mapping loss value. S6. Repeat S1~S5 until all loss values ​​and values ​​converge to obtain the final detection model; S7. Obtain the image of the drone to be detected, process it using only a fixed resolution, and input it into the final detection model for detection, outputting the drone category and candidate box coordinates.

2. The method according to claim 1, characterized in that, S2 includes: In the feature fusion module of the detection model, the random scale feature map and the fixed scale feature map are scale-aligned to obtain an aligned feature map; The aligned feature map is multiplied element-wise with the fixed-scale feature map to obtain the modulation feature map; The modulation feature map is superimposed element-wise with the fixed-scale feature map to obtain a preliminary enhanced feature map; The initial enhanced feature map is subjected to ReLU activation function operation and then batch normalized to obtain the enhanced fixed-scale feature map.

3. The method according to claim 1, characterized in that, S3 includes: The enhanced fixed-scale feature map and random-scale feature map are used to predict target candidate boxes, their coordinates, and confidence scores using the coarse detection head of the detection model. The target candidate boxes are then filtered according to a preset confidence threshold to remove background candidate boxes with confidence scores below the threshold, resulting in fixed-scale candidate boxes and random-scale candidate boxes. Calculate the scale conversion factor based on the ratio of fixed resolution to random resolution, and map the fixed-scale candidate box to the random-scale feature map based on the scale conversion factor to obtain the mapped candidate box. Boundary verification is performed on the mapped candidate boxes, and mapped candidate boxes that exceed the range of the random scale feature map are removed; Calculate the intersection-union ratio (IU) between the filtered mapped candidate boxes and the random-scaled candidate boxes, and retain the mapped candidate boxes whose IU is greater than or equal to a preset IU threshold; The candidate box mapping loss value is calculated based on the retained mapping candidate boxes and the random-scale candidate boxes.

4. The method according to claim 3, characterized in that, The candidate box mapping loss value is calculated according to the following formula: Lroi = α×L_reg+β×L_iou; When |GT-Pred|≤1, L_reg=0.5×(GT-Pred)²; when |GT-Pred|>1, L_reg=|GT-Pred|-0.5; L_iou = -ln(IOU); Wherein, α is the first preset weight coefficient, β is the second preset weight coefficient, L_reg is the first coordinate regression loss value, L_iou is the intersection-union ratio loss value, Lroi is the candidate box mapping loss value, GT is the coordinates of the retained mapped candidate box, Pred is the coordinates of the random scale candidate box; and IOU is the mean of the intersection-union ratio of the retained mapped candidate box and the random scale candidate box.

5. The method according to claim 3, characterized in that, S4 includes: Based on the coordinates of the retained mapping candidate boxes and the random scale candidate boxes, corresponding local feature maps are cropped from the fixed scale feature map and the random scale feature map after alignment with the fixed scale feature map. The cropped local feature map is input into the local super-resolution module for feature-level super-resolution enhancement to obtain a mapped local enhanced feature map and a random-scale local enhanced feature map; the local super-resolution module includes a channel attention submodule and a residual learning submodule; Noise filtering is applied to the mapped local augmentation feature map and the random-scale local augmentation feature map; The filtered mapped local augmented feature map and the filtered random scale local augmented feature map are subjected to pseudo-texture avoidance to obtain the mapped super-resolution local feature map and the random scale super-resolution local feature map.

6. The method according to claim 5, characterized in that, S5 includes: Align the fixed-scale candidate boxes with the mapped super-resolution local feature maps and the random-scale super-resolution local feature maps to obtain the aligned fixed-scale candidate boxes and the aligned mapped super-resolution local feature maps and the random-scale super-resolution local feature maps. The aligned fixed-scale candidate boxes, the aligned mapped super-resolution local feature maps, and the random-scale super-resolution local feature maps are flattened to obtain the fixed-scale feature vector, the mapped super-resolution local feature vector, and the random-scale super-resolution local feature vector. The semantic consistency loss value is calculated by using the cosine similarity loss function to obtain the semantic consistency loss value from the fixed-scale feature vector, the local feature vector after mapping super-resolution, and the local feature vector after random-scale super-resolution. The aligned mapped super-resolution local feature map and the aligned random scale super-resolution local feature map are input into the fine detection head of the detection model to predict the target candidate box, its coordinates, category, and category confidence. Based on the predicted target candidate box, its coordinates, category, and category confidence, as well as the target calibration box, its coordinates, and category of the manually calibrated original UAV image, the category loss value and the second coordinate regression loss value are calculated. The target loss value is calculated based on the semantic consistency loss value, candidate box mapping loss value, category loss value, and second coordinate regression loss value; the parameters of the detection model are then updated in reverse based on the target loss value.

7. A progressive unmanned aerial vehicle (UAV) detection system based on super-resolution guidance, characterized in that, include: The feature extraction unit is used to acquire the original UAV image, process it with fixed resolution and random resolution respectively, and input it into the backbone network of the detection model to extract fixed-scale feature maps and random-scale feature maps respectively. The feature enhancement unit is used to perform scale alignment and feature interaction between the random scale feature map and the fixed scale feature map to generate an enhanced fixed scale feature map. The coarse prediction unit is used to predict fixed-scale candidate boxes and random-scale candidate boxes based on the enhanced fixed-scale feature map and random-scale feature map, respectively, and to map the fixed-scale candidate boxes to the random-scale feature map. The candidate box mapping loss value is calculated based on the mapping result and the random-scale candidate boxes. The super-resolution enhancement unit is used to crop local feature maps from fixed-scale feature maps and aligned random-scale feature maps based on the mapped candidate boxes and random-scale candidate boxes, and then perform super-resolution enhancement on them to obtain two super-resolution local feature maps. The reverse update unit is used to calculate the semantic consistency loss value of the fixed-scale candidate boxes and two super-resolution local feature maps according to the cosine similarity loss function, and update the parameters of the detection model by combining the candidate box mapping loss value; The repeated training unit is used to repeat the feature extraction unit, feature enhancement unit, coarse prediction unit, super-resolution enhancement unit, and reverse update unit until all loss values ​​and values ​​converge to obtain the final detection model. The detection unit is used to acquire images of the drones to be detected, processes them using only a fixed resolution, and inputs them into the final detection model for detection, outputting the drone category and candidate box coordinates.

8. The system according to claim 7, characterized in that, The feature enhancement unit includes: The alignment subunit is used to perform scale alignment between the random scale feature map and the fixed scale feature map in the feature fusion module of the detection model to obtain an aligned feature map. The dot product sub-unit is used to perform element-wise dot product between the aligned feature map and the fixed-scale feature map to obtain the modulation feature map; The superposition subunit is used to superimpose the modulation feature map and the fixed-scale feature map element by element to obtain a preliminary enhanced feature map; The enhancement subunit is used to perform ReLU activation function operation on the preliminary enhanced feature map and obtain the enhanced fixed-scale feature map through batch normalization.

9. The system according to claim 7, characterized in that, The coarse prediction unit includes: The filtering subunit is used to predict target candidate boxes, their coordinates, and confidence scores by passing the enhanced fixed-scale feature map and random-scale feature map through the coarse detection head of the detection model; and to filter the target candidate boxes according to a preset confidence threshold to remove background candidate boxes with confidence scores below the threshold, thereby obtaining fixed-scale candidate boxes and random-scale candidate boxes. The mapping subunit is used to calculate the scale conversion factor based on the ratio of fixed resolution and random resolution, and to map the fixed-scale candidate box to the random-scale feature map according to the scale conversion factor to obtain the mapped candidate box. The elimination sub-unit is used to perform boundary verification on the mapping candidate boxes and eliminate mapping candidate boxes that exceed the range of the random scale feature map. The retention sub-unit is used to calculate the intersection-union ratio (IU) between the filtered mapped candidate boxes and the random-scaled candidate boxes, and to retain the mapped candidate boxes whose IU is greater than or equal to a preset IU threshold. The computation subunit is used to calculate the candidate box mapping loss value based on the retained mapping candidate boxes and random-scale candidate boxes.

10. The system according to claim 9, characterized in that, The candidate box mapping loss value is calculated according to the following formula: Lroi = α×L_reg+β×L_iou; When |GT-Pred|≤1, L_reg=0.5×(GT-Pred)²; when |GT-Pred|>1, L_reg=|GT-Pred|-0.5; L_iou = -ln(IOU); Wherein, α is the first preset weight coefficient, β is the second preset weight coefficient, L_reg is the first coordinate regression loss value, L_iou is the intersection-union ratio loss value, Lroi is the candidate box mapping loss value, GT is the coordinates of the retained mapped candidate box, Pred is the coordinates of the random scale candidate box; and IOU is the mean of the intersection-union ratio of the retained mapped candidate box and the random scale candidate box.

Images