Feature fusion method and device under large-scale small target detection, equipment and medium

Abstract

The embodiment of the invention provides a feature fusion method and device under large-scale small target detection, equipment and a medium, and the method comprises the steps: obtaining image features in a traffic scene, carrying out the processing of the image features, carrying out the extraction of target key features through a longitudinal and transverse mechanism, and obtaining a feature map; performing feature fusion on the feature map based on a multi-dimensional feature fusion technology; recombining the target prediction frame by adopting a target detection optimization technology so as to reinforce the feature fusion model; small targets which are missed in the convolution process are captured through a vertical and horizontal mechanism and feature correlation, richer and more accurate features are provided for feature fusion, the problem that upper and lower layer features are not aligned is solved through a multi-dimensional feature fusion technology, targets which are missed and have key features lost are captured through a target detection optimization technology, and the target detection accuracy is improved. The feature fusion model is enhanced, and the resolution, accuracy and efficiency of target detection are improved.

Description

Technical Field

[0001] The present application relates to the field of feature fusion technology, and specifically, to a feature fusion method, device, equipment and medium for large-scale small target detection. Background Art

[0002] With the rise of deep convolutional neural networks, general object detection has made significant progress. These general object detection algorithms can effectively detect and locate objects of various sizes, benefiting many practical application scenarios. However, despite the prosperity of the general object detection field, the field of small object detection has not made significant progress as expected. Small object detection (SOD) is a subfield of object detection that focuses on detecting objects with smaller volumes. It has important theoretical and practical significance in various scenarios such as surveillance, drone scene analysis, pedestrian detection, and traffic sign detection in autonomous driving. For small object detection, there is a lack of large-scale datasets tailored for it. In addition, due to the small size of the target, low resolution of the target, and possible overlap between targets, traditional target detection algorithms are difficult to meet the requirements of accuracy and efficiency, resulting in the relatively lagging development of the field of small object detection. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide a feature fusion method, device, equipment and medium for large-scale small target detection, so as to solve the problems of low resolution, low accuracy and low efficiency of existing traditional target detection algorithms.

[0004] In a first aspect, an embodiment of the present application provides a feature fusion method for large-scale small target detection, including:

[0005] Obtain image features in traffic scenes, process the image features, and extract target key features through vertical and horizontal mechanisms to obtain feature maps;

[0006] Perform feature fusion on feature maps based on multi-dimensional feature fusion technology;

[0007] The target detection optimization technique is used to reorganize the target prediction boxes to strengthen the feature fusion model.

[0008] In the above implementation process, the image features in the traffic scene are obtained, the image features are processed, and the key features of the target are extracted through the vertical and horizontal mechanisms to obtain a feature map; the feature map is fused based on the multi-dimensional feature fusion technology; the target detection optimization technology is used to reorganize the target prediction frame to strengthen the feature fusion model; through the vertical and horizontal mechanisms, the feature correlation is used to capture small targets that are omitted in the convolution process, providing richer and more accurate features for feature fusion, and the multi-dimensional feature fusion technology is used to solve the problem of misalignment of the upper and lower layer features. The target detection optimization technology is used to capture the targets that are omitted or lose key features, strengthen the feature fusion model, and improve the resolution, accuracy and efficiency of target detection.

[0009] Furthermore, the feature fusion of the feature graph based on the multi-dimensional feature fusion technology includes:

[0010] Define the target key features of the feature graph according to preset rules;

[0011] Identify and emphasize key areas in the image through the spatial perception module to focus on key spatial features;

[0012] Fusion of features of different scales based on scale-aware module;

[0013] A channel-aware module is used to conduct competition and cooperation between channels.

[0014] In the above implementation process, the model is made more sensitive and accurate in terms of space, scale and channel, which significantly improves the ability to accurately identify unknown targets. By combining multi-dimensional optimization of space, scale and channel features, the model can handle target detection tasks in complex scenarios more efficiently.

[0015] Furthermore, the target key features of the feature graph are defined according to preset rules, including:

[0016] Given the concatenation of features from L different levels in the feature pyramid:

[0017]

[0018] Use upsampling or downsampling to adjust the size of the features of the successive levels to make them consistent with the proportions of the features of the intermediate levels;

[0019] Represent the rescaled feature pyramid as a 4D tensor:

[0020] X∈R H×W×L×C ;

[0021] Where L represents the number of layers in the pyramid, H, W, and C represent the height, width, and number of channels of the intermediate layer features, respectively;

[0022] definition:

[0023] S = H × W;

[0024] Reshape the tensor into a three-dimensional tensor:

[0025] X∈R S×L×C ;

[0026] Using the general formula of self-attention:

[0027] W(X) = F(X)·X;

[0028] Among them, F(X) is an attention function;

[0029] Convert the attention function into three sequential attentions, each focusing on only one view:

[0030] W(f)=F C (F L ·F S (X)·X)·X;

[0031] Among them, F S (X), F L and F C (F L ·F S (X)·X) are three different attention functions applied to dimensions S, L, and C, respectively.

[0032] In the above implementation process, by shaping four-dimensional tensors and three-dimensional tensors according to the feature pyramid, the role of each tensor dimension can be explored, and the three-dimensional attention function can be obtained according to the three-dimensional tensor, and feature fusion can be performed through subsequent perception calculations.

[0033] Furthermore, the key areas in the image are identified and emphasized by the spatial perception module to focus on key spatial features, including:

[0034] Use deformable convolution to sparse spatial perception learning so that it focuses on a set spatial location;

[0035] Adjust the position of the convolution kernel through deformable convolution so that the convolution kernel captures the salient area in the feature map;

[0036] The formula for cross-level feature aggregation is performed at the same spatial position captured by the convolution kernel:

[0037]

[0038] Where k is the number of sparse sampling positions; position p k The spatial offset is moved to p by self-learning k +Δpk ; Δm k is the position p k The self-learning importance scalar at .

[0039] In the above implementation, the feature map becomes more sparse and focused on the salient spatial locations of foreground objects through the spatial perception module; this module enables the model to better focus on important spatial features by identifying and emphasizing key areas in the image.

[0040] Furthermore, the fusing of features of different scales based on the scale perception module includes:

[0041] By analyzing the semantic information of features, the importance of features at different scales can be determined;

[0042] The weights of features at each scale are adjusted according to the importance of features at different scales, so that the scale-aware model can capture and fuse multi-scale information. The formula is as follows:

[0043]

[0044] in, is a linear function approximated by a 1×1 convolutional layer, and σ(x)=max(0, min(1, (x+1) / 2)) is an activation function.

[0045] In the above implementation process, after being processed by the scale-aware module, the model becomes more sensitive to the scale differences of foreground objects; the scale-aware mechanism dynamically fuses features of different scales according to the semantic importance of the features, thereby enhancing the model's target detection capabilities at different scales.

[0046] Furthermore, the channel perception module is used to perform competition and cooperation between channels, including:

[0047] Aggregate the global context information in each channel based on the global context embedding module;

[0048] A normalization method is used to create a competitive relationship between channels;

[0049] Competition and cooperation between different channels are promoted through gating mechanisms.

[0050] In the above implementation process, the channel perception module promotes competition and cooperation between channels. This mechanism improves the model's expression and selectivity of channel information by performing frequency analysis and gating adjustment on channel characteristics, thereby improving the ability to recognize unknown targets.

[0051] Further, the aggregating global context information in each channel based on the global context embedding module includes:

[0052] The two-dimensional discrete cosine transform was introduced, and multiple two-dimensional DCT frequency components were used to compress information;

[0053] Split the input features into multiple parts along the channel dimension to exploit frequency characteristics;

[0054] The input features [F0, F1, ..., F n-1 ] is expressed as:

[0055] F i ∈R C×H×W ,i∈{0,1,…,n-1},

[0056] Where C' is the channel compression, and C is divisible by n;

[0057] For each part, the corresponding two-dimensional discrete cosine transform DCT frequency component is assigned, and the result of the DCT frequency component is used as the channel-focused compression result. The formula is as follows:

[0058]

[0059] Where X i [u i ,v i ] is the two-dimensional index of the frequency component, i∈{0,1,…,n-1}, Freq i ∈R C "Compressed C" dimensional vector;

[0060] Given embedding weights α = [α1,…,α C ], the model formula is defined as:

[0061]

[0062] Among them, Freq is the compression vector of the entire channel, Freq∈R C is a multi-spectral vector; ε = 1e5 is a constant, using a trainable parameter α C To control the weight of each channel.

[0063] In the above implementation, the global context information in each channel is aggregated, and this module can develop global context information outside the small receptive field of the convolutional layer.

[0064] Furthermore, the use of a normalization method to create a competitive relationship between channels includes:

[0065] A normalization method is used to create a competitive relationship between neurons;

[0066] Let S = [S1,…S C ], the channel normalization formula is:

[0067]

[0068] Among them, ε = 1e5 is a constant, scalar For S C The scale is normalized.

[0069] In the above implementation process, a competitive relationship between neurons (or channels) can be created, with lightweight computing resources and stable training performance.

[0070] Furthermore, the gating mechanism is used to promote competition and cooperation between different channels, including:

[0071] Let the gate weight γ = [γ1,…γ C ] and the gate bias β = [β1,…β C ], the gating function is:

[0072]

[0073] The scale of the initial channel is adjusted through the gating function; when the gating weight of the channel is positively activated, the competition between the channel and other channels is promoted;

[0074] When the gating weight of a channel is negatively activated, the channel is encouraged to cooperate with other channels;

[0075] When the gating weight and gating bias are zero, the gating function allows the original feature to pass through the next layer, and F(X) = 1;

[0076] Initialize γ and β to 0 when initializing the layer, and the generation formula is:

[0077]

[0078] In the above implementation process, a gating mechanism is used to adapt to the original features; by introducing the initial channel of the gating mechanism, competition and cooperation can be promoted during the training process.

[0079] Furthermore, the target detection optimization technology is used to reorganize the target prediction frame to strengthen the feature fusion model, including:

[0080] Generate a prediction box for each pixel of the image feature;

[0081] Get the boundary set of the predicted box, calculate the difference between the boundary set and the boundary of the real box, and sort the boundary set of the predicted box according to the corresponding rank;

[0082] The boundaries of a set of sets with the same rank are reorganized into a new prediction box, and the IOU overlap between the new prediction box and the real box boundary is calculated.

[0083] In the above implementation process, the positive and negative samples are redistributed so that the positive samples are more suitable for the real frame and the negative samples have more background information. From the perspective of target detection optimization, the targets that are missed or have lost key features are captured to further enhance the detection performance of the model.

[0084] In a second aspect, an embodiment of the present application provides a feature fusion device for large-scale small target detection, including:

[0085] The feature acquisition module is used to acquire the image features in the traffic scene, process the image features, and extract the key features of the target through the vertical and horizontal mechanisms to obtain the feature map;

[0086] A feature fusion module is used to fuse feature maps based on multi-dimensional feature fusion technology;

[0087] The feature enhancement module is used to reorganize the target prediction box using target detection optimization technology to enhance the feature fusion model.

[0088] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0089] A processor, a memory and a bus, wherein the processor is connected to the memory via the bus, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, they are used to implement the feature fusion method for large-scale small target detection as described above.

[0090] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a server, implements the feature fusion method for large-scale small target detection as described above.

[0091] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes instructions, and when the instructions are executed by a computer, the computer implements the method as described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0093] Figure 1 A flowchart of a feature fusion method for large-scale small target detection provided in an embodiment of the present application;

[0094] Figure 2 A schematic diagram of a box reorganization algorithm for a feature fusion method for large-scale small target detection provided in an embodiment of the present application;

[0095] Figure 3 A flowchart of another feature fusion method for large-scale small target detection provided by an embodiment of the present application;

[0096] Figure 4 A schematic diagram of target scale AP value comparison of corresponding modules of a feature fusion method for large-scale small target detection provided in an embodiment of the present application;

[0097] Figure 5 A schematic diagram showing comparison of AP values ​​of nine categories of corresponding modules of a feature fusion method for large-scale small target detection provided in an embodiment of the present application;

[0098] Figure 6 A system function display diagram of a feature fusion method for large-scale small target detection provided by an embodiment of the present application;

[0099] Figure 7 A detection effect diagram of VFNet provided in an embodiment of the present application;

[0100] Figure 8 A detection effect diagram of a proposed algorithm for a feature fusion method for large-scale small target detection provided in an embodiment of the present application;

[0101] Fig. 9 It is a structural schematic diagram of a feature fusion device for large-scale small target detection provided by an embodiment of the present application;

[0102] Fig.10 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0103] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0104] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0105] There is a lack of large-scale datasets tailored for small target detection. In addition, due to the small size of the target, low resolution of the target, and possible overlap between targets, traditional target detection algorithms are difficult to meet the requirements of accuracy and efficiency, resulting in the relatively lagging development of the field of small target detection, which has not yet reached the ideal level. Among them, selecting a suitable large-scale small target detection dataset and designing a suitable network structure and training strategy can improve the accuracy and efficiency of small target detection, thereby effectively coping with the challenges brought by large-scale small target detection.

[0106] The training of general target detection algorithms usually uses large-scale general data sets, such as COCO, PASCAL VOC, etc., which mainly contain medium and large-sized targets. However, for small targets, the number of samples in these data sets is small, and they are usually unevenly distributed in the image. The lack of enough small target samples will make it difficult for small target detection algorithms to learn and generalize, limiting their performance improvement. Therefore, in order to promote the development of the field of small target detection, it is necessary to establish a large-scale data set specifically for small targets. The two existing large-scale small target detection data sets for driving and aviation scenes: SODA-D and SODA-A, are intended to promote the development of the field of small target detection. Based on the existing work, the embodiment of the present application adopts an improved VFNe framework.

[0107] In traffic scenes, there are a large number of small targets such as pedestrians, bicycles, and electric vehicles. They are relatively small in size and move fast, which can easily lead to traffic accidents. In order to improve traffic safety, it is necessary to detect these small targets in a timely manner. VFNet performs well in dense target detection. Therefore, using VFNet for large-scale small target detection can help drivers be more vigilant and prevent traffic accidents. It can also assist traffic supervision departments in real-time monitoring of traffic violations, taking law enforcement measures in a timely manner, and maintaining traffic order and safety. This is very important for traffic safety, traffic supervision and law enforcement, driving assistance, and urban benefits.

[0108] VFNet is an anchor-free detection network based on FCOS and ATSS, which adopts the standard paradigm of anchor-free network: backbone+FPN+decoupling head. VFNet believes that the center branch cannot accurately reflect the regression characteristics of the target, so it removes the center branch used in FCOS and ATSS and adds three components: VFL loss function, star regression box, and bounding box fine-tuning module. VFL is improved from Focal Loss. Compared with Focal Loss, VFL uses asymmetric parameters to weight the positive and negative samples of the detector, so that the weight distribution of the positive and negative samples of the detector is easier to fit the number distribution of positive and negative samples in the data set; the star-shaped bounding box is based on deformable convolution, which captures targets of different shapes and sizes, and enhances the detector's detection ability for deformed targets; the bounding box fine-tuning module dynamically adjusts the bounding box by learning the scaling factors of the four sides of the bounding box, so that the bounding box fits the real box better. In addition, VFNet also incorporates the bounding box score in the prediction process into the classification branch, generating the (Iou-awareCls_score) IACS score, combining the features of classification and regression, and providing richer features for classification tasks. Thanks to its more reasonable feature adjustment and joint feature expression, VFNet has become the mainstream model in the anchor-free network model, and has excellent detection effects on targets of various scales. Due to the removal of the restrictions of the anchor box, when detecting large-scale small targets, the model will not be overloaded due to a large number of negative samples, and the detection performance will not deviate from the negative samples, which is very important for the detection of large-scale small targets.

[0109] Although VFNet has completed the detection of targets of different scales under traffic background with the advantage of anchor-free network, due to the limitations of its own network structure design, VFNet still has a lot of room for improvement. First of all, the FPN structure in VFNet includes "multi-scale detection" and "feature fusion", but YOLOF points out that feature fusion plays a very small role in the FPN structure. Therefore, the embodiment of the present application redesigns the FPN structure to strengthen its feature fusion from the aspects of lateral connection, spatial and scale dynamic fusion, etc. Secondly, the free bounding box regression of the anchor-free network is its advantage and also its limitation. The free bounding box regression allows the network to detect more objects of different sizes and shapes, which perfectly meets the needs of SODA-D, but it also causes it to be unable to accurately regress the target box. In the case of NMS, many valid boxes are suppressed, resulting in a small amount of missed detection. Therefore, the embodiment of the present application starts with feature fusion and bounding box regression, adds a new network structure to VFNet, and designs a new model LSOD-TS.

[0110] Based on the above problems, the present application embodiment proposes a feature fusion method for large-scale small target detection.

[0111] Please see Figure 1 , Figure 1 A flow chart of a feature fusion method for large-scale small target detection provided in an embodiment of the present application. The feature fusion method for large-scale small target detection includes:

[0112] 100. Obtain image features in a traffic scene, process the image features, and extract target key features through a vertical and horizontal mechanism to obtain a feature map.

[0113] Specifically, to address the problem of inaccurate position perception, a horizontal and vertical mechanism (SHV) is proposed, which uses strip convolution and pooling in the horizontal and vertical directions to aggregate global features, achieves long-distance modeling of features with a very small number of parameters, and strengthens the correlation between features, so that small targets with random distribution and poor discriminability can be enhanced, allowing the network to more accurately capture the key features of the target and alleviate the problem of small target positioning and missed detection. It can be understood that the embodiment of the present application does not specifically limit the process of extracting key features of the target through the horizontal and vertical mechanism.

[0114] Optionally, the image features in the embodiment of the present application may be image features obtained by processing pictures or videos of traffic scenes.

[0115] As mentioned above, the model is guided to focus on important areas in the image and suppress unnecessary noise information through the vertical and horizontal mechanisms.

[0116] 200. Feature fusion is performed on feature maps based on multi-dimensional feature fusion technology.

[0117] It can be understood that multidimensional feature fusion technology is an advanced feature fusion method, which can dynamically select and integrate feature information from different levels or scales according to the specific characteristics of the input data and task requirements. This technology dynamically adjusts the fusion strategy to make the model more flexible in processing features of different positions and scales, thereby improving the prediction accuracy and performance of the model.

[0118] The embodiment of the present application optimizes the efficiency of feature utilization and improves detection performance through multi-dimensional feature fusion technology. By dynamically fusing features of different scales, the model can more accurately locate the target object and reduce false detections and missed detections.

[0119] 300. The target detection optimization technology is used to reorganize the target prediction box to strengthen the feature fusion model.

[0120] Specifically, a prediction box is generated for each pixel of the image feature; the boundary set of the prediction box is obtained, the difference between the boundary set and the boundary of the real box is calculated, and the boundary set of the prediction box is sorted according to the corresponding rank; the boundaries of a set of sets with the same rank are reorganized into a new prediction box, and the IOU overlap between the new prediction box and the boundary of the real box is calculated; thereby, the positive and negative samples are redistributed, so that the positive samples are more fitted to the real box, and the negative samples have more background information, so as to capture the targets that are omitted or lose key features from the perspective of target detection optimization, and further enhance the detection performance of the model.

[0121] It is understandable that compared with the anchor-based network, the anchor-free network has two contributions in small targets: first, it can detect objects of more shapes and scales. The small targets of SODA-D are not ordinary fixed-size targets. In the traffic background, not only are there many types of objects, but they also change shape. As traffic moves, the scale of objects also changes. The anchor-free network is more suitable for multi-scale and multi-deformation small target detection. Secondly, the number of anchor frames of the anchor-free network is less than that of the anchor-based network. For small targets, more valid anchor frames will be saved when performing NMS.

[0122] The anchor-free network also has natural limitations. The anchor frame regression is too free and cannot accurately fit the target. In the case of NMS, many valid frames are suppressed, resulting in a small number of missed detections. To solve this problem, the embodiment of the present application proposes a box reorganization algorithm. After releasing the anchor frame, the box reorganization method is used to select high-quality positive and negative samples to better fit the real frame, thereby achieving the purpose of reducing missed detections.

[0123] For example, please refer to Figure 2 , generate a prediction box p for each pixel i of image I i = {l i ,r i ,b i ,t i}, the set of predicted boxes is denoted as B I ={p0,p1,…,p n}. Where l, t, r, b represent the left, top, right, and bottom boundaries respectively. To represent the real box, the four categories of boundaries are divided into four sets, represented as: left I ={l0,l1,…,l n}, right I ={r0,r1,…,r n}, bottom I ={b0,bl1,…,b n}, top I ={t0,t1,…,t n}.

[0124] For the boundary set of the predicted box, calculate the boundary with the real box The difference The bounding box sets of the prediction boxes are then sorted according to the corresponding ranks.

[0125] After sorting, the boundaries of the four sets with the same rank are reorganized into a new box: B' I ={p'0,p1',…,p' n}, and then reorganize the prediction box p i 'With the real frame Perform IOU overlap calculation.

[0126] In the above implementation process, the image features in the traffic scene are obtained, the image features are processed, and the key features of the target are extracted through the vertical and horizontal mechanisms to obtain a feature map; the feature map is fused based on the multi-dimensional feature fusion technology; the target detection optimization technology is used to reorganize the target prediction frame to strengthen the feature fusion model; through the vertical and horizontal mechanisms, the feature correlation is used to capture small targets that are omitted in the convolution process, providing richer and more accurate features for feature fusion, and the multi-dimensional feature fusion technology is used to solve the problem of misalignment of the upper and lower layer features. The target detection optimization technology is used to capture the targets that are omitted or lose key features, strengthen the feature fusion model, and improve the resolution, accuracy and efficiency of target detection.

[0127] Based on the above embodiments, Figure 3 A flow chart of another feature fusion method for large-scale small target detection provided by an embodiment of the present application is given. The feature fusion method for large-scale small target detection is a specific embodiment of the feature fusion method for large-scale small target detection. Figure 3 ,The feature fusion method for large-scale small target detection includes:

[0128] 210. Define the target key features of the feature map according to preset rules.

[0129] Given the concatenation of features from L different levels in the feature pyramid:

[0130]

[0131] Use upsampling or downsampling to adjust the size of the features of the successive levels to make them consistent with the proportions of the features of the intermediate levels;

[0132] Represent the rescaled feature pyramid as a 4D tensor:

[0133] X∈R H×W×L×C ;

[0134] Where L represents the number of layers in the pyramid, H, W, and C represent the height, width, and number of channels of the intermediate layer features, respectively;

[0135] definition:

[0136] S = H × W;

[0137] Reshape the tensor into a three-dimensional tensor:

[0138] X∈R S×L×C ;

[0139] Among them, there is a certain correlation between the geometric transformation of different object shapes and the features of different spatial positions. Therefore, by improving the feature learning algorithm, the target detection model's perception of spatial information can be improved; secondly, there is a correlation between the scale difference of the target and the features of different levels. By improving the feature learning methods at different levels, the target detection model's perception of targets of different scales can be enhanced; finally, the representation of different targets may be related to the features of different channels. Therefore, improving the representation learning of different channel features can improve the target detection model's perception of channel information.

[0140] Specifically, the general formula of self-attention is adopted:

[0141] W(X) = F(X)·X;

[0142] Among them, F(X) is an attention function;

[0143] Convert the attention function into three sequential attentions, each focusing on only one view:

[0144] W(f)=F C (F L ·F S (X)·X)·X;

[0145] Among them, F S (X), F L and F C (F L ·F S (X)·X) are three different attention functions applied to dimensions S, L, and C, respectively.

[0146] As mentioned above, by shaping four-dimensional tensors and three-dimensional tensors according to the feature pyramid, we can explore the role of each tensor dimension, obtain the three-dimensional attention function according to the three-dimensional tensor, and perform feature fusion through subsequent perceptual calculations.

[0147] 220. The key areas in the image are identified and emphasized through the spatial perception module to focus on key spatial features.

[0148] Understandably, in order to further improve spatial perception, feature aggregation F is introduced to focus on finding common discriminative regions between different spatial locations and feature layers. S . Considering the high-dimensional feature representation S, this module is divided into two steps: first, deformable convolution is used to sparse the attention learning so that it can focus on a specific spatial position. Deformable convolution can flexibly adjust the position of the convolution kernel so that it can better capture the salient areas in the feature map, thereby enhancing the perception ability of the model in the spatial dimension; next, cross-level feature aggregation is performed at these same spatial positions to promote information fusion and sharing. Through cross-level feature aggregation, features at different levels can complement and strengthen each other, thereby forming a richer and more discriminative feature representation. This process can effectively integrate feature information of different scales and levels, and enhance the model's discrimination ability in complex scenes. As a result, the spatial perception ability of the model is significantly enhanced, making it more accurate and efficient in processing complex images and identifying different targets.

[0149] Specifically, deformable convolution is used to sparse spatial perception learning so that it can be focused on a set spatial position; the position of the convolution kernel is adjusted through deformable convolution so that the convolution kernel captures the salient area in the feature map; and cross-level feature aggregation formula is performed at the same spatial position captured by the convolution kernel:

[0150]

[0151] Where k is the number of sparse sampling positions; position p k The spatial offset is moved to p by self-learning k +Δp k ; Δm k is the position p k The self-learning importance scalar at .

[0152] As mentioned above, through the spatial perception module, the feature map becomes more sparse and focuses on the significant spatial positions of foreground objects; this module enables the model to better focus on important spatial features by identifying and emphasizing key areas in the image.

[0153] 230. Features of different scales are fused based on the scale perception module.

[0154] It can be understood that the scale-aware mechanism dynamically fuses features of different scales according to their semantic importance.

[0155] Specifically, by analyzing the semantic information of the features, the importance of the features at different scales is determined; the weights of the features at each scale are adjusted according to the importance of the features at different scales, so that the scale-aware model can capture and fuse multi-scale information. The formula is as follows:

[0156]

[0157] in, is a linear function approximated by a 1×1 convolutional layer, and σ(x)=max(0, min(1, (x+1) / 2)) is an activation function.

[0158] Therefore, after being processed by the scale-aware module, the model becomes more sensitive to the scale differences of foreground objects; the scale-aware mechanism dynamically fuses features of different scales according to the semantic importance of the features, thereby enhancing the model's target detection capability at different scales.

[0159] 240. A channel perception module is used to conduct competition and cooperation between channels.

[0160] 241. Aggregate the global context information in each channel based on the global context embedding module.

[0161] Among them, information with a large receptive field helps to avoid local blurring caused by information with a small receptive field (e.g., convolutional layers). Therefore, a global context embedding module is designed to aggregate the global context information in each channel, which can exploit the global context information beyond the small receptive field of the convolutional layer.

[0162] In the channel attention mechanism, the use of intervals results in retaining only the lowest frequency information, while other frequency components are discarded, which encode useful information patterns when representing the channel and should not be ignored. In order to better compress channel information and introduce more information, a two-dimensional discrete cosine transform DCT frequency component is introduced, and multiple two-dimensional DCT frequency components are used to compress information; the input feature F is split into multiple parts along the channel dimension to utilize frequency features;

[0163] The input features [F0, F1, ..., F n-1 ] is expressed as:

[0164] F i ∈R C×H×W ,i∈{0,1,…,n-1},

[0165] Wherein, C' is channel compression, and C is divisible by n; channel compression is set to reduce the number of channels to reduce operation consumption. For example, if the number of channels is 256, the channels are compressed by 256 / 16=16.

[0166] For each part, the corresponding two-dimensional discrete cosine transform DCT frequency component is assigned, and the result of the DCT frequency component is used as the channel-focused compression result. The formula is as follows:

[0167]

[0168] Where X i [u i ,v i ] is the two-dimensional index of the frequency component, i∈{0,1,…,n-1}, Freq i ∈R C "Compressed C" dimensional vector;

[0169] Given embedding weights α = [α1,…,α C ], the model formula is defined as:

[0170]

[0171] Among them, Freq is the compression vector of the entire channel, Freq∈R C is a multi-spectral vector; where ε = 1e5 is a constant to avoid the problem of derivative at zero point; in addition, since different channels have different meanings, a trainable parameter α is used C To control the weight of each channel;

[0172] If α C Close to 0, channel normalization is independent of the corresponding channel i, that is, the gating weight α C Let the initial channel F C Ability to learn the unique features of a channel to distinguish other channels.

[0173] Thus, the global context information in each channel is aggregated, and this module can exploit the global context information outside the small receptive field of the convolutional layer.

[0174] 242. A normalization method is used to create a competitive relationship between channels.

[0175] Specifically, the normalization method can create a competitive relationship between neurons (or channels), with lightweight computing resources and stable training performance. Similar to LRN (Local Response Normalization), an L2 normalization is used to operate across channels, i.e. channel normalization.

[0176] Let S = [S1,…S C ], the channel normalization formula is:

[0177]

[0178] Among them, ε = 1e5 is a constant, scalar For S C The scale is normalized to avoid S when C is large. C The scale is too small.

[0179] In the above implementation process, a competitive relationship between neurons (or channels) can be created, with lightweight computing resources and stable training performance.

[0180] 243. Promote competition and cooperation among different channels through gating mechanisms.

[0181] Specifically, gated adaptive adjustment adopts a gating mechanism, namely gated adaptation, to adapt to the original features. By introducing the gating mechanism, F C It can promote competition and cooperation during the training process. Let the gate weight γ = [γ1,…γ C ] and the gate bias β = [β1,…β C ], the gating function is:

[0182]

[0183] Each initial channel F C The scale of will be adjusted by its corresponding gate, i.e., F(X). Since channel normalization is parameter-free, trainable weights γ and gating biases β are introduced to flexibly control the activation of gated channels. Traditional local response normalization (LRN) relies only on competition between neurons, while the channel-aware module, by combining normalization methods and gating mechanisms, is able to model various relationships (including competition and cooperation) between different channels. When the gating weight γ of a channel C When being activated, F C Similar to LRN, it promotes competition between the channel and other channels. On the contrary, when the gating weight is negatively activated, the channel-aware module encourages the channel to cooperate with other channels. This dual mechanism enables the model to dynamically adjust the expression of channel features in different situations, thereby more effectively capturing and utilizing the relationship between channels and improving the performance of the model.

[0184] When the gating weight and gating bias are zero, the gating function allows the original feature to pass through the next layer, and F(X) = 1;

[0185] Initialize γ and β to 0 when initializing the layer, and the generation formula is:

[0186]

[0187] It is understandable that a gating mechanism is used to adapt to the original features; by introducing the initial channel of the gating mechanism, competition and cooperation can be promoted during the training process.

[0188] Therefore, through the role of the channel perception module, competition and cooperation between channels are promoted. This mechanism improves the model's expression and selectivity of channel information by performing frequency analysis and gating adjustment on channel characteristics, thereby improving the ability to recognize unknown targets.

[0189] As described above, the embodiments of the present application make the model more sensitive and accurate in terms of space, scale and channel, significantly improving the ability to accurately identify unknown targets. By combining multi-dimensional optimization of space, scale and channel features, the model can more efficiently handle target detection tasks in complex scenarios.

[0190] Through the spatial perception module, the feature map becomes sparser and focuses on the significant spatial location of the foreground object. This module enables the model to better focus on important spatial features by identifying and emphasizing key areas in the image. Next, after being processed by the scale perception module, the model becomes more sensitive to the scale differences of foreground objects. The scale perception mechanism dynamically fuses features of different scales according to the semantic importance of the features, thereby enhancing the model's ability to detect objects at different scales. Finally, through the role of the channel perception module, competition and cooperation between channels are promoted. This mechanism improves the model's ability to express and select channel information by performing frequency analysis and gating adjustments on channel features, thereby improving the ability to recognize unknown targets. This optimized design makes the model more sensitive and accurate in three aspects: space, scale, and channel, and significantly improves the ability to accurately recognize unknown targets. By combining multi-dimensional optimization of spatial, scale, and channel features, the model can more efficiently handle object detection tasks in complex scenes.

[0191] The embodiment of the present application designs a horizontal and vertical mechanism (SHV, Spatial attention based on Horizontal and Vertical convolution) in the horizontal connection, realizes long-distance modeling with a very small amount of parameters, and uses feature correlation to capture small targets that are omitted in the convolution process, providing richer and more accurate features for feature fusion; secondly, in the feature fusion process, a multi-dimensional feature fusion module (UFNET, Unified Feature Network) is designed, and deformable convolution is used to solve the problem of misalignment of upper and lower layer features, and then the key features of the target are emphasized through channel attention, and then the feature interaction is completed from the two perspectives of frequency domain components and channel gating mechanisms, so that the feature map after feature fusion has a reasonable weight distribution in both space and channels. Finally, on the distributor of ATSS, the box reorganization algorithm is used to redistribute positive and negative samples, so that the positive samples are more fitted to the real box, and the negative samples have more background information, and the targets that are omitted and lose key features are captured from the perspective of the box, further enhancing the detection performance of the model.

[0192] For example, Figure 4 The comparison of AP values ​​of target scales of the corresponding modules of the embodiment of the present application is shown, demonstrating the effectiveness of VFnet and the three algorithms. mAP represents the mean of AP under multiple thresholds, AP50 represents the accuracy when the IOU threshold is 0.5, AP75 represents the accuracy when the IOU threshold is 0.75, APes, APrs, APgs, and APN distribution represent the detection accuracy of extremely small, relatively small, generally small, and normal target scales in the data set. It can be seen that among the three algorithms, SHV has the largest average index improvement, and it is superior to other algorithms in many indicators, among which APes has the best effect. This is mainly because SHV can achieve effective long-distance modeling, and through interaction with global features, it enhances the key features of small targets that are not discriminative and easily suppressed, so that even extremely small targets can be detected, which is of great significance for the detection of small targets in traffic scenes. UFNET does not improve very small targets much, but it improves normal targets significantly. This is mainly because in the process of feature fusion, larger targets can obtain more critical and discriminative features. Small targets lack such discriminative features and are easily suppressed. In the BOX reconstruction algorithm, the real box with weak discriminability is fitted from the perspective of the box, making the box regression more accurate, and the improvement in mAP and AP50 is more obvious. When the three algorithms are combined together, there is no conflict between the three, but they promote each other. In the final algorithm, a significant improvement has been achieved in average accuracy and various scales, which fully proves the effectiveness of the algorithm proposed in the embodiment of the present application.

[0193] For example, Figure 5 The figure shows the comparison of AP values ​​of 9 categories of the corresponding modules of the embodiment of the present application. It can be seen that each algorithm basically improves VFnet. The final algorithm also combines the advantages of the three algorithms and achieves the best in most single categories, which fully demonstrates the effectiveness of the algorithm proposed in the embodiment of the present application.

[0194] Also, please refer to Figure 6 The system functions of the embodiments of the present application include: (1) Mode selection: three modes are provided: image detection, video detection, and real-time detection. (2) Model selection: five models are provided: vfnet_baseline, vfnet_UFnet, vfnet_SHV, vfnet_box, and vfnet_all. (3) Image / video file selection: select an image / video for detection. (4) Image / video detection result display: in order to facilitate viewing of images and observe the changes before and after running the model, the system provides a visual interface for images, which allows you to more intuitively view the display effect of the image after detection. The display styles provided are: original image and processed visual effect image.

[0195] Figure 7 The image shown is the detection effect of VFNet. Figure 8 The displayed image is the detection effect of the algorithm proposed in the embodiment of the present application. By comparison, the algorithm proposed in the embodiment of the present application can not only detect more small targets and solve the problem of missed detection, but also improve the confidence of the detected targets, thereby improving the target detection performance of the model in traffic scenarios from many aspects, and plays a vital role in real life.

[0196] The above steps are not to be performed in a strict order as described in the numbers, but should be understood as an overall solution.

[0197] In the second aspect, based on the above embodiments, Fig. 9 This is a schematic diagram of the structure of a feature fusion device for large-scale small target detection provided by an embodiment of the present application. Fig. 9 The feature fusion device for large-scale small target detection provided in this embodiment specifically includes: a feature acquisition module 901, a feature fusion module 902 and a feature enhancement module 903.

[0198] Among them, the feature acquisition module 901 is used to obtain image features in traffic scenes, process the image features, and extract target key features through vertical and horizontal mechanisms to obtain feature maps; the feature fusion module 902 is used to fuse feature maps based on multi-dimensional feature fusion technology; the feature enhancement module 903 is used to reorganize the target prediction frame using target detection optimization technology to enhance the feature fusion model.

[0199] As described above, the embodiment of the present application obtains image features in a traffic scene, processes the image features, and extracts target key features through a vertical and horizontal mechanism to obtain a feature map; performs feature fusion on the feature map based on a multi-dimensional feature fusion technology; uses a target detection optimization technology to reorganize the target prediction frame to strengthen the feature fusion model; uses a vertical and horizontal mechanism to utilize feature correlation to capture small targets that are omitted in the convolution process, providing richer and more accurate features for feature fusion, solves the problem of misalignment of upper and lower layer features through a multi-dimensional feature fusion technology, uses a target detection optimization technology to capture targets that are omitted or have lost key features, strengthens the feature fusion model, and improves the resolution, accuracy, and efficiency of target detection.

[0200] The feature fusion device for large-scale small target detection provided in the embodiment of the present application can be used to execute the feature fusion method for large-scale small target detection provided in the above embodiment, and has corresponding functions and beneficial effects.

[0201] In a third aspect, an embodiment of the present application further provides an electronic device that can integrate the feature fusion device for large-scale small target detection provided by an embodiment of the present application. Fig.10 Schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Fig.10 The electronic device includes: an input device 73, an output device 74, a memory 72 and one or more processors 71; the memory 72 is used to store one or more programs; when the one or more programs are executed by the one or more processors 41, the one or more processors 71 implement the feature fusion method for large-scale small target detection provided by the above embodiment. The input device 73, the output device 74, the memory 72 and the processor 71 can be connected by a bus or other means. Fig.10 The example of connecting through bus is taken in the following.

[0202] The processor 71 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 72, that is, realizes the above-mentioned feature fusion method under large-scale small target detection.

[0203] The electronic device provided above can be used to execute the feature fusion method for large-scale small target detection provided in the above embodiment, and has corresponding functions and beneficial effects.

[0204] In a fourth aspect, an embodiment of the present application also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the feature fusion method for large-scale small target detection as described above, and can achieve the same beneficial effects.

[0205] Of course, the storage medium containing computer executable instructions provided in an embodiment of the present application is not limited to the feature fusion method under large-scale small target detection as described above, and can also execute related operations in the feature fusion method under large-scale small target detection provided in any embodiment of the present application.

[0206] In a fifth aspect, the embodiments of the present application also provide a computer program product, and the methods described in the various embodiments of the present application can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instruction is loaded and executed on a computer, the process or function described in the various embodiments of the present application is executed in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user device, a core network device, an OAM (Open Application Model) or other programmable device.

[0207] The computer program or instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program or instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired or wireless means. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; it may also be an optical medium, such as a digital video disk; it may also be a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0208] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0209] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0210] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for enabling an electronic device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0211] The above description is only an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0212] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0213] It should be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements limited by the sentence "comprise one..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements.

Claims

1. A feature fusion method for large-scale small target detection, characterized in that: include: Obtain image features in traffic scenes, process the image features, and extract target key features through vertical and horizontal mechanisms to obtain feature maps; Perform feature fusion on feature maps based on multi-dimensional feature fusion technology; The target detection optimization technique is used to reorganize the target prediction boxes to strengthen the feature fusion model.

2. The feature fusion method for large-scale small target detection according to claim 1 is characterized in that: The method of fusing the feature graph based on the multi-dimensional feature fusion technology includes: Define the target key features of the feature graph according to preset rules; Identify and emphasize key areas in the image through the spatial perception module to focus on key spatial features; Fusion of features of different scales based on scale-aware module; A channel-aware module is used to conduct competition and cooperation between channels.

3. The feature fusion method for large-scale small target detection according to claim 2 is characterized in that: Defining the target key features of the feature graph according to the preset rules includes: Given the concatenation of features from L different levels in the feature pyramid: Use upsampling or downsampling to adjust the size of the features of the successive levels to make them consistent with the proportions of the features of the intermediate levels; Represent the rescaled feature pyramid as a 4D tensor: X∈R H×W×L×C ; Where L represents the number of layers in the pyramid, H, W, and C represent the height, width, and number of channels of the intermediate layer features, respectively; definition: S = H × W; Reshape the tensor into a three-dimensional tensor: X∈R S×L×C ; Using the general formula of self-attention: W(X) = F(X)·X; Among them, F(X) is an attention function; Convert the attention function into three sequential attentions, each focusing on only one view: W(f)=F C (F L ·F S (X)·X)·X; Among them, F S (X), F L and F C (F L ·F S (X)·X) are three different attention functions applied to dimensions S, L, and C, respectively.

4. The feature fusion method for large-scale small target detection according to claim 2 is characterized in that: The spatial perception module identifies and emphasizes key areas in the image to focus on key spatial features, including: Use deformable convolution to sparse spatial perception learning so that it focuses on a set spatial location; Adjust the position of the convolution kernel through deformable convolution so that the convolution kernel captures the salient area in the feature map; The formula for cross-level feature aggregation is performed at the same spatial position captured by the convolution kernel: Where k is the number of sparse sampling positions; position p k The spatial offset is moved to p by self-learning k +Δp k ; Δm k is the position p k The self-learning importance scalar at .

5. The feature fusion method for large-scale small target detection according to claim 2, characterized in that: The fusing of features of different scales based on the scale perception module includes: By analyzing the semantic information of features, the importance of features at different scales can be determined; The weights of features at each scale are adjusted according to the importance of features at different scales, so that the scale-aware model can capture and fuse multi-scale information. The formula is as follows: in, is a linear function approximated by a 1×1 convolutional layer, and σ(x)=max(0, min(1, (x+1) / 2)) is an activation function.

6. The feature fusion method for large-scale small target detection according to claim 2, characterized in that: The channel sensing module is used to perform competition and cooperation between channels, including: Aggregate the global context information in each channel based on the global context embedding module; A normalization method is used to create a competitive relationship between channels; Competition and cooperation between different channels are promoted through gating mechanisms.

7. The feature fusion method for large-scale small target detection according to claim 6, characterized in that: The aggregating global context information in each channel based on the global context embedding module includes: Two-dimensional discrete cosine transform (DCT) frequency components are introduced, and multiple two-dimensional DCT frequency components are used to compress information; Split the input features into multiple parts along the channel dimension to exploit frequency characteristics; The input features [F0, F1, ..., F n-1 ] is expressed as: Where C' is the channel compression, and C is divisible by n; For each part, the corresponding two-dimensional discrete cosine transform DCT frequency component is assigned, and the result of the DCT frequency component is used as the channel-focused compression result. The formula is as follows: Where X i [u i ,v i ] is the two-dimensional index of the frequency component, i∈{0,1,…,n-1}, Freq i ∈R C "Compressed C" dimensional vector; Given embedding weights α = [α1,…,α C ], the model formula is defined as: Among them, Freq is the compression vector of the entire channel, Freq∈R C is a multi-spectral vector; ε = 1e5 is a constant, using a trainable parameter α C To control the weight of each channel.

8. The feature fusion method for large-scale small target detection according to claim 7, characterized in that: The normalization method is used to create a competitive relationship between channels, including: A normalization method is used to create a competitive relationship between neurons; Let S = [S1,…S C ], the channel normalization formula is: Among them, ε = 1e5 is a constant, scalar For S C The scale is normalized.

9. The feature fusion method for large-scale small target detection according to claim 8, characterized in that: The gating mechanism promotes competition and cooperation between different channels, including: Let the gate weight γ = [γ1,…γ C ] and the gate bias β = [β1,…β C ], the gating function is: The scale of the initial channel is adjusted through the gating function; when the gating weight of the channel is positively activated, the competition between the channel and other channels is promoted; When the gating weight of a channel is negatively activated, the channel is encouraged to cooperate with other channels; When the gating weight and gating bias are zero, the gating function allows the original feature to pass through the next layer, and F(X) = 1; When initializing the layer, γ and β are initialized to 0, and the generation formula is:

10. The feature fusion method for large-scale small target detection according to claim 1, characterized in that: The target detection optimization technology is used to reorganize the target prediction frame to strengthen the feature fusion model, including: Generate a prediction box for each pixel of the image feature; Get the boundary set of the predicted box, calculate the difference between the boundary set and the boundary of the real box, and sort the boundary set of the predicted box according to the corresponding rank; The boundaries of a set of sets with the same rank are reorganized into a new prediction box, and the IOU overlap between the new prediction box and the real box boundary is calculated.

11. A feature fusion device for large-scale small target detection, characterized in that: include: The feature acquisition module is used to acquire the image features in the traffic scene, process the image features, and extract the key features of the target through the vertical and horizontal mechanisms to obtain the feature map; A feature fusion module is used to fuse feature maps based on multi-dimensional feature fusion technology; The feature enhancement module is used to reorganize the target prediction box using target detection optimization technology to enhance the feature fusion model.

12. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the processor is connected to the memory via the bus, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, they are used to implement the feature fusion method for large-scale small target detection as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a server, implements the feature fusion method for large-scale small target detection as described in any one of claims 1-10.

14. A computer program product, characterized in that The computer program product comprises instructions which, when executed by a computer, cause the computer to implement the method according to any one of claims 1-10.

Images