Small target detection method based on multi-scale super-resolution and adaptive fusion

Abstract

The invention belongs to the technical field of computer vision and intelligent detection, and discloses a small target detection method based on multi-scale super-resolution and adaptive fusion. The method comprises the following steps: extracting a multi-level feature map from an input image through a backbone network, processing a shallow feature map through a pixel rearrangement technology and a detail attention generation mechanism, and connecting the shallow feature map with an original shallow feature residual error to obtain an enhanced shallow feature map; calculating the attention weight of each level of feature map through global average pooling and a multilayer perceptron, carrying out normalization, and carrying out weighted fusion on each level of feature map by using the normalized attention weight; and inputting the fused feature map into a detection head based on a Transform architecture to obtain a target category and bounding box prediction, determining a confidence threshold according to the ratio of the bounding box area to the total area of the input image, and obtaining a small target detection result according to the confidence threshold. According to the method, the calculation overhead is effectively controlled while the model performance is remarkably improved, and the dual requirements for precision and efficiency in practical application are met.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent detection technology, specifically relating to a small target detection method based on multi-scale super-resolution and adaptive fusion. Background Technology

[0002] In recent years, the demand for small target detection in applications such as infrastructure security, industrial quality inspection, and remote sensing monitoring has grown rapidly. Because small targets occupy a very small proportion of an image, have a low signal-to-noise ratio, and are often elongated in shape (such as cracks, scratches, and micropores), they are prone to losing details and edge information during conventional downsampling and feature extraction processes. Scale variations and background interference further exacerbate the detection difficulty, leading to high false negatives and false negatives in existing systems for small target scenarios. Meanwhile, practical engineering requirements demand a balance between detection accuracy and real-time performance, placing higher demands on model structure and computational efficiency.

[0003] Most existing technologies rely on convolutional neural networks or the Transformer paradigm for feature extraction and multi-scale fusion. Traditional top-down or lateral fusion methods such as FPN (Feature Pyramid Network) and PAFPN (Path Aggregation Feature Pyramid Network) often employ fixed or simple weighted fusion strategies, making it difficult to achieve adaptive fine-grained selection for targets at different locations and scales. This results in insufficient semantics for small targets, blurred edges, and inadequate background suppression. While end-to-end detection methods based on DETR (Detection Transformer) and RT-DETR (Real-Time Detection Transformer) have the advantage of global modeling, their shallow backbone layers also lose crucial high-frequency details and texture information during large-step downsampling. Furthermore, most methods are still insufficient in detail enhancement, exhibiting weak responses to fine-grained textures and edges, making it difficult to stably represent small, elongated, and low-contrast targets. On the other hand, directly stacking complex attention or reconstruction modules will result in significant computational and memory overhead, which is not conducive to engineering deployment and real-time detection. Summary of the Invention

[0004] The purpose of this invention is to provide a small target detection method based on multi-scale super-resolution and adaptive fusion. This method significantly improves model performance while effectively controlling computational overhead, meeting the dual requirements of accuracy and efficiency in practical applications.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A small target detection method based on multi-scale super-resolution and adaptive fusion includes:

[0007] The system receives an input image, extracts multi-level feature maps through a backbone network, inputs the shallow feature maps from the multi-level feature maps into the super-resolution feature reconstruction module, processes them through pixel rearrangement technology and detail attention generation mechanism and connects them with the original shallow feature residuals to obtain the enhanced shallow feature map.

[0008] Attention weights for each layer of feature maps are calculated and normalized using global average pooling and multilayer perceptron. The normalized attention weights are then used to perform weighted fusion of the feature maps at each layer to obtain a fused feature map.

[0009] The fused feature map is input into the detection head based on the Transformer architecture to obtain the target category and bounding box prediction. The confidence threshold is determined based on the ratio of the bounding box area to the total area of ​​the input image, and the target categories and bounding boxes with a confidence score greater than the confidence threshold are retained as small target detection results.

[0010] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0011] Preferably, the received input image is processed by a backbone network to extract multi-level feature maps, including:

[0012] Receive the input image and perform preprocessing and standardization on it;

[0013] The preprocessed and standardized input image is processed using a backbone network to output a preliminary multi-level feature map.

[0014] Align the channel dimensions of each level of feature map using 1×1 convolution to obtain the final multi-level feature map.

[0015] Preferably, the pixel rearrangement technique involves the following process:

[0016] The channel dimension of the shallow feature map is expanded by a factor of 4 using a 1×1 convolution;

[0017] The information from the channel dimension, which has been expanded by 4 times, is rearranged into the spatial dimension through pixel rearrangement, achieving a 2x upsampling, and the processing is complete.

[0018] Preferably, the detail attention generation mechanism processes the following steps:

[0019] A single-layer 1×1 convolution is used to compress the channels of the shallow feature map into a single channel, and an activation function is used to process the single-channel shallow feature map to obtain a detail attention map.

[0020] Preferably, the enhanced shallow feature map is obtained by concatenating it with the original shallow feature residuals, including:

[0021] Element-wise multiplication is performed on the output of the pixel rearrangement technique and the output of the detail attention generation mechanism. The multiplication result is then passed through a 3×3 convolution and residually connected to the original shallow features to obtain the enhanced shallow feature map.

[0022] Preferably, the detection head based on the Transformer architecture includes a multi-layer Transformer encoder, a multi-layer Transformer decoder, and a prediction head. The prediction head includes a classification head for predicting the probability distribution of the target category and a regression head for predicting the normalized bounding box coordinates.

[0023] Preferably, determining the confidence threshold based on the ratio of the bounding box area to the total area of ​​the input image includes:

[0024]

[0025] in, The determined confidence threshold, Based on the confidence threshold, For adjustment coefficients, For decay rate control parameters, This is the ratio of the bounding box area to the total area of ​​the input image.

[0026] As a preferred method, the loss function during the training phase is calculated as follows:

[0027]

[0028] in, For loss function, For the set of small target samples in the training batch, This refers to the sample number of the small target. Indicates the first The normalized value of the ratio of the bounding box area of ​​each small target sample to the total area of ​​the input image. Indicates the first Classification loss for small target samples, Indicates the first The regression loss for a small target sample.

[0029] The present invention provides a small target detection method based on multi-scale super-resolution and adaptive fusion, which has the following advantages compared with the prior art: (1) It proposes a small target detection framework based on multi-scale super-resolution and adaptive fusion, including a super-resolution feature reconstruction module, adaptive multi-scale fusion and Transformer query detection head, to achieve end-to-end detection, weaken the dependence of anchor boxes and NMS, and improve detection accuracy and efficiency. (2) By introducing a super-resolution feature reconstruction module to restore shallow details and cooperating with scale attention weighted fusion, the contribution of each scale is adaptively allocated according to the content, which significantly strengthens the edge and texture representation of small targets, and replaces the traditional multi-layer spatial attention module of convolution or dense connection with a lightweight scale attention mechanism, which reduces the computational overhead while maintaining detection performance. (3) It adopts a dynamic confidence threshold based on the relative area of ​​the target and works in conjunction with the small target weighted loss to adaptively control the prediction according to the scale in the output stage, taking into account both recall and precision, and effectively improving the recall rate of small target detection. Attached Figure Description

[0030] Figure 1 This is a flowchart of the small target detection method based on multi-scale super-resolution and adaptive fusion of the present invention.

[0031] Figure 2 This is a flowchart of the small target object feature enhancement process of the present invention;

[0032] Figure 3 This is a schematic diagram of the super-resolution feature reconstruction module of the present invention;

[0033] Figure 4 This is a schematic diagram of the structure of the multi-level feature fusion module of the present invention; Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0036] To address the problems of high false negative rates, difficulty in balancing detail and semantics in current small object detection methods, weak response to low-contrast targets, and insufficient efficiency of multi-scale fusion, this invention proposes a small object detection method based on multi-scale super-resolution and adaptive fusion. This method constructs a super-resolution feature reconstruction module in the shallow layers of the backbone network, effectively recovering edge and texture features lost during downsampling through pixel rearrangement upsampling and detail attention mechanisms, thus enhancing the saliency of small objects. A scale-aware attention fusion mechanism is employed, generating weights for each scale through global average pooling and MLP (Multilayer Perceptron), achieving alignment and weighted fusion of multi-scale features, and outputting semantically sufficient and detail-preserving fused features. A query-based detection head enables end-to-end prediction, avoiding anchor box design and NMS (Non-Maximum Suppression) post-processing. During the training phase, a weighted composite loss for small objects is used to improve convergence efficiency, and during the inference phase, a dynamic confidence threshold is set based on the relative area of ​​the target to balance recall and precision. This method significantly improves the recall and positioning accuracy of small target detection while effectively controlling computational overhead, meeting the dual requirements of accuracy and efficiency in practical applications. It is suitable for detecting small targets such as cracks, scratches, and micropore defects in scenarios such as infrastructure security, industrial quality inspection, and remote sensing monitoring.

[0037] like Figure 1 As shown in the figure, a small target detection method based on multi-scale super-resolution and adaptive fusion in this embodiment specifically includes the following steps:

[0038] (1) Receive the input image, extract multi-level feature maps through the backbone network, input the shallow feature maps in the multi-level feature maps into the super-resolution feature reconstruction module, process them through pixel rearrangement technology and detail attention generation mechanism and connect them with the original shallow feature residuals to obtain the enhanced shallow feature maps.

[0039] like Figure 2 As shown, in this embodiment, multi-level feature maps are extracted via a backbone network. Define the channel dimension and spatial resolution of features at different scales, and based on this, analyze shallow features. , Super-resolution reconstruction is performed, and pixel rearrangement upsampling and detail attention generation mechanisms are used to recover the edge and texture loss caused by downsampling, resulting in an enhanced shallow representation.

[0040] (1-1) Input preprocessing and standardization of original images:

[0041] The input images are analyzed, with typical original resolutions including 1280×720, 1920×1080 (video / surveillance), 4032×3024, 4000×3000 (mobile / camera), and aerial images larger than 4000×3000 (often sliced ​​to 1024×1024 or 2048×2048 during training). To ensure batch processing consistency, the above input images are scaled proportionally and padded to a standard size S×S (default S=640), followed by channel-level zero-mean, unit-variance normalization. Pixel values ​​are then mapped to a zero-mean, unit-variance distribution.

[0042]

[0043] in The original image matrix is ​​given by: and This normalization is a constant (e.g., RGB three-channel) pre-statistically obtained on the training set, which remains unchanged during training and inference. This normalization is insensitive to the input image resolution, so images of different resolutions can share the same set. and This method is effective in preventing the vanishing gradient problem in deep networks and accelerating model convergence, provided that the preprocessing steps (color space, scaling / filling method) are consistent.

[0044] (1-2) Construction of multi-scale feature pyramid:

[0045] After preprocessing the input image, the standardized image data enters the feature extraction stage. In order to preserve both fine-grained and high-level semantic information, this embodiment uses a mature backbone network (such as ResNet) to construct a feature pyramid, extracting feature layers with four different receptive fields from the input image. The corresponding downsampling step sizes are respectively Among them, shallow features and It retains rich low-level visual information such as edges and textures, which is crucial for the accurate localization of small targets; deep features and It encodes high-level semantics and global context, providing key clues for target category discrimination.

[0046]

[0047] To facilitate subsequent multi-scale fusion operations, the backbone network is recorded as follows: The original characteristics of the stage output are ( The step sizes are respectively This is denoted as the initial multi-level feature map. To facilitate multi-scale fusion, a 1×1 convolution is used. Align its channels to a uniform dimension ,get This operation only changes the number of channels, preserving the [previous / previous] channel count. Achieve the same spatial resolution (Hk×Wk) A unified representation. This not only reduces computational complexity but also allows features at different levels to be directly subjected to arithmetic operations, laying the foundation for subsequent adaptive fusion.

[0048] (1-3) Constructing a super-resolution feature reconstruction module:

[0049] like Figure 3 As shown, although the backbone network successfully extracts multi-scale feature maps, shallow feature maps inevitably lose crucial details during downsampling, which is particularly detrimental to the detection of small targets. This embodiment addresses this by constructing a lightweight super-resolution feature reconstruction module specifically for shallow feature maps. and To restore details and avoid the loss of edge texture information, the specific processing flow is as follows:

[0050] (1-3-1) Improving the spatial resolution of feature maps through pixel rearrangement techniques: Here, we first use 1×1 convolution. Expand the channel dimension of the shallow feature map by a factor of 4 to C×4, and then perform pixel rearrangement operations. These channel-dimensional information are rearranged into the spatial dimension, achieving a 2x upsampling, resulting in the output of the pixel rearrangement technique, denoted as the feature. Compared to traditional bilinear interpolation, pixel rearrangement can learn a more refined upsampling pattern, effectively recovering high-frequency details lost during downsampling.

[0051]

[0052] (1-3-2) Introduce a lightweight attention mechanism to highlight important details:

[0053]

[0054] in Using the Sigmoid activation function, the output is a detail attention map in the range [0,1]. A single-layer 1×1 convolution is used. By compressing the channels of the C-dimensional shallow feature map to a single channel, the number of parameters is low, making it easy to train. It can be replaced with a two-layer bottleneck structure to further enhance the representation. This lightweight design allows the super-resolution feature module to increase computational overhead by only a small amount (less than 5%). Depending on the actual hardware conditions, it can be selectively applied only to… Apply this module, or simultaneously apply it to and It flexibly balances performance and efficiency, ensuring that this attention map can adaptively identify and enhance local features such as edges and corners that are crucial for small object detection.

[0055] (1-3-3) The enhanced features are fused with the original features through residual connections:

[0056]

[0057] in, For the enhanced ⊙ denotes element-wise multiplication. 3×3 convolution. Residual connections are used to smooth out artifacts that may occur during the fusion process, while residual connections ensure that the original semantic information is not lost during enhancement. After feature reconstruction, the enhanced feature set is output. The details of the shallow features are significantly restored, which can provide richer feature representations for subsequent small object detection and provide preliminary detailed features for subsequent detection heads, ensuring the ability to focus on and integrate them in the later stages.

[0058] (2) such as Figure 4 As shown, this embodiment constructs a scale-aware fusion module to align the resolution of multi-scale features. Attention weights for each level of feature map are calculated and normalized using global average pooling and a multilayer perceptron. These normalized attention weights are then used to perform weighted fusion of the feature maps at each level, resulting in a semantically sufficient and detail-preserving fused feature map.

[0059] (2-1) Scale attention calculation for different image scenes:

[0060] The key to multi-scale feature fusion lies in dynamically determining the contribution of features at each layer: scenes containing many small targets rely more on shallow detail features, while scenes dominated by large targets require deeper semantic features. To address this, this embodiment designs an adaptive scale attention mechanism: first, global average pooling is used to extract global statistics for each layer's features to characterize the overall activation pattern of that layer; then, a lightweight multilayer perceptron is used to map the global statistics to importance weights. Finally, to ensure the weights are comparable and interpretable, the Softmax function is used to normalize the weights at each scale, making their sum equal to 1, thus intuitively representing the contribution ratio of each layer to the fusion result. This allows the model to adaptively adjust the fusion strategy according to the input content, improving its adaptability to different scenes.

[0061] In this embodiment, each layer of feature map (original deep feature map or enhanced shallow feature map, denoted as) is first processed. Apply global average pooling This is done to obtain global statistics that represent the overall activation pattern of the layer; these statistics are then input into a lightweight MLP consisting of two fully connected layers to obtain the first... Attention weights corresponding to hierarchical feature maps The first layer is used for channel compression (e.g., compressing to one-sixteenth of the original) followed by ReLU activation; the second layer maps the compressed representation to a single scalar fraction, representing the unnormalized importance of that scale. This is denoted as:

[0062]

[0063]

[0064] To ensure comparability of weights across different scales and that their sum equals 1, the scores are softmax normalized along the scale dimension. This yields the normalized attention weights. And satisfy This indicates that the result has been normalized along the scale dimension. It is then used for weighted fusion after feature alignment, thereby achieving an adaptive balance between shallow details and deep semantics.

[0065] (2-2) Adaptive weighted fusion of attention:

[0066] After calculating the attention weights for each scale, the next step is to apply these weights to the actual feature fusion process. After obtaining the scale attention weights, weighted fusion of multi-scale features is performed. Since the features at each layer have different spatial resolutions, they first need to be aligned to a uniform target scale through interpolation. This embodiment selects... The resolution (step size 8) is used as the fusion scale, balancing spatial detail preservation and computational efficiency.

[0067]

[0068] in, For the first Feature maps at various scales Indicates will Align to the target fusion scale (default in this embodiment) The required scaling factor (for the space dimensions); This indicates a two-dimensional interpolation alignment operation (using bilinear or nearest neighbor; this embodiment defaults to bilinear), which will... Resampling to Same height and width. The above summation is an element-wise weighted summation: that is, first sum the features at each scale through... Align to the same resolution, then adjust the scalar weights. The weighted fusion is expanded to H×W×C via broadcasting, and then multiplied element-wise with the aligned features before summing. Vectorized tensor operations are used in the implementation instead of explicit loops. This weighted fusion avoids complex cascading or iterative structures, while leveraging learned weights... Content-adaptive feature integration was achieved. The fused features... It also contains information from various scales: shallow detailed textures, mid-level local structures, and deep semantic context, providing a comprehensive and balanced feature representation for subsequent object detection.

[0069] (3) Input the fused feature map into the detection head (encoder-decoder detection head) based on the Transformer architecture to obtain the target category and bounding box prediction. Determine the confidence threshold based on the ratio of the bounding box area to the total area of ​​the input image, and retain the target category and bounding box with a confidence greater than the confidence threshold as the small target detection result.

[0070] This embodiment integrates feature inputs with a detection head based on the Transformer architecture, and uses a query mechanism to interact with features to directly generate target category and bounding box predictions. During the training phase, a composite supervision of classification loss, localization loss, and small target weighted loss is used to improve the convergence efficiency of small targets. During the inference phase, a dynamic confidence threshold is set based on the relative area of ​​the target to control the output quality and effectively balance recall and precision.

[0071] (3-1) Transformer-based detection head: includes a multi-layer Transformer encoder, a multi-layer Transformer decoder and a prediction head, wherein the prediction head includes a classification head for predicting the probability distribution of the target class and a regression head for predicting the normalized bounding box coordinates.

[0072] Thanks to the high-quality detailed features provided by the super-resolution path feature reconstruction module, the fused features The system then employs a Transformer-based detection head that directly predicts targets using a query mechanism. Unlike traditional anchor-box-based detectors, this embodiment uses a set of learnable target queries to represent potential target instances. Each query vector is responsible for detecting a target in the image, and its representation is progressively refined through interaction with image features. The query update process is implemented through a multi-layer Transformer decoder.

[0073]

[0074]

[0075]

[0076] in, Indicates the first The query set of the layer, Indicates the first The query set of the layer, For cross-attention modules, It refers to layer-by-layer iterative computation Next, for the first There are layers The transformation ensures that the key / value memory remains the same fused feature throughout the entire decoding process. (Keep this constant) to ensure that queries can focus on relevant regions on the feature map. Then, through an attention mechanism, each query learns to focus on the spatial location and feature patterns of the target it is responsible for detecting. Feedforward networks further process the attention output, enhancing the non-linear expressive power of features.

[0077] After After iterative optimization of the layers, each query converges to a stable representation, and then the final prediction is generated through classification and regression heads: the classification head outputs the class probability distribution. , and These are the weight matrix and bias vector of the classification head, and the normalized bounding box predicted by the regression head. , For bounding box coordinates, The width of the bounding box. This represents the height of the bounding box. Where... This refers to the feature vector of a single query output by the last layer of the decoder; that is, the representation of a query after updates across the attention and feedforward networks. Each query has a corresponding... This is used to feed the classification head into the system to generate the class probabilities and bounding box parameters for the query. The classification head and regression head share the same input: the former outputs... The latter outputs bounding box parameters. The implementation uses an activation function to constrain the coordinates to a normalized range of 0–1 (e.g., the [0,1] image coordinate system). This end-to-end prediction method avoids the anchor box design and non-maximum suppression (NMS) post-processing in traditional detectors, greatly simplifying the detection process.

[0078] (3-2) Dynamic confidence threshold optimizes output quality:

[0079] The raw prediction results generated by the detection head need to be optimized to obtain the final detection results. This step directly determines the final performance of the detector. For small target detection tasks, output optimization is particularly critical because it directly affects whether small targets can be successfully detected. Under a fixed threshold mechanism, a large number of real small targets are incorrectly filtered out because their confidence is slightly below the threshold, leading to a sharp drop in recall. However, simply lowering the global threshold introduces a large number of false detections, damaging precision. Therefore, this embodiment proposes a dynamic confidence threshold strategy based on the relative area of ​​the target to achieve differentiated output control for targets of different scales. The core idea is to adaptively adjust the confidence threshold according to the relative size of the detected target:

[0080]

[0081] in, The determined confidence threshold, The base confidence threshold (usually set around 0.5). This is the adjustment coefficient (controlling the adjustment range of the threshold). This is a decay rate control parameter (which determines the sensitivity of the threshold to changes in target size). It is the ratio of the bounding box area to the total area of ​​the input image (reflecting the relative scale of the target).

[0082] When the target size is extremely small ( When the exponential term approaches zero, the threshold automatically drops to a region significantly below the base threshold (approximately). ), significantly increasing the recall opportunity for small targets; as the target size increases ( As the threshold increases, the exponential term approaches 1, and the threshold asymptotically converges to the baseline threshold level. This maintains strict selection criteria for major targets. It should be noted that these parameters are not fixed constraints; adjustment coefficients are also possible. and decay rate control parameters It can be flexibly adjusted according to specific needs such as the target scale distribution characteristics and false detection tolerance of the actual dataset. This design provides a reliable default configuration while retaining the flexibility to adapt to different application scenarios.

[0083] The method of this invention updates the parameters of each network during the training phase, and the calculated loss function is as follows:

[0084]

[0085] in, For loss function, For the set of small target samples in the training batch, This refers to the sample number of the small target. Indicates the first The normalized value of the ratio of the bounding box area of ​​each small target sample to the total area of ​​the input image. Indicates the first Classification loss for small target samples, Indicates the first The regression loss is calculated for a small number of target samples. This loss function is achieved through... This approach implements inverse frequency weighting, applying stronger supervisory signals to smaller targets. This allows the model to prioritize optimizing the feature representations of smaller targets during training, systematically improving their baseline confidence level. This training strategy, together with the dynamic threshold mechanism in the inference phase, forms a cohesive optimization system that ultimately significantly improves the recall performance of small targets while maintaining high accuracy, fundamentally solving the core challenge in small target detection.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A small target detection method based on multi-scale super-resolution and adaptive fusion, characterized in that, The small target detection method based on multi-scale super-resolution and adaptive fusion includes: The system receives an input image, extracts multi-level feature maps through a backbone network, inputs the shallow feature maps from the multi-level feature maps into the super-resolution feature reconstruction module, processes them through pixel rearrangement technology and detail attention generation mechanism and connects them with the original shallow feature residuals to obtain the enhanced shallow feature map. Attention weights for each layer of feature maps are calculated and normalized using global average pooling and multilayer perceptron. The normalized attention weights are then used to perform weighted fusion of the feature maps at each layer to obtain a fused feature map. The fused feature map is input into the detection head based on the Transformer architecture to obtain the target category and bounding box prediction. The confidence threshold is determined based on the ratio of the bounding box area to the total area of ​​the input image, and the target categories and bounding boxes with a confidence score greater than the confidence threshold are retained as small target detection results.

2. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The received input image is processed by a backbone network to extract multi-level feature maps, including: Receive the input image and perform preprocessing and standardization on it; The preprocessed and standardized input image is processed using a backbone network to output a preliminary multi-level feature map. Align the channel dimensions of each level of feature map using 1×1 convolution to obtain the final multi-level feature map.

3. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The pixel rearrangement technique involves the following process: The channel dimension of the shallow feature map is expanded by a factor of 4 using a 1×1 convolution; The information from the channel dimension, which has been expanded by 4 times, is rearranged into the spatial dimension through pixel rearrangement, achieving a 2x upsampling, and the processing is complete.

4. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The detail attention generation mechanism processes the following steps: A single-layer 1×1 convolution is used to compress the channels of the shallow feature map into a single channel, and an activation function is used to process the single-channel shallow feature map to obtain a detail attention map.

5. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The enhanced shallow feature map is obtained by concatenating it with the original shallow feature residuals, including: Element-wise multiplication is performed on the output of the pixel rearrangement technique and the output of the detail attention generation mechanism. The multiplication result is then passed through a 3×3 convolution and residually connected to the original shallow features to obtain the enhanced shallow feature map.

6. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The detection head based on the Transformer architecture includes a multi-layer Transformer encoder, a multi-layer Transformer decoder, and a prediction head. The prediction head includes a classification head for predicting the probability distribution of the target class and a regression head for predicting the normalized bounding box coordinates.

7. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The step of determining the confidence threshold based on the ratio of the bounding box area to the total area of ​​the input image includes: in, The determined confidence threshold, Based on the confidence threshold, For adjustment coefficients, For decay rate control parameters, This is the ratio of the bounding box area to the total area of ​​the input image.

8. The small target detection method based on multi-scale super-resolution and adaptive fusion according to claim 1, characterized in that, The loss function during the training phase is calculated as follows: in, For loss function, For the set of small target samples in the training batch, This refers to the sample number of the small target. Indicates the first The normalized value of the ratio of the bounding box area of ​​each small target sample to the total area of ​​the input image. Indicates the first Classification loss for small target samples, Indicates the first The regression loss for a small target sample.

Images