An unmanned aerial vehicle target detection method based on feature fusion DINO

Abstract

The application discloses a kind of unmanned aerial vehicle target detection methods based on feature fusion DINO, belong to target detection technical field. Including the following steps: obtaining original unmanned aerial vehicle image and pre-processing;The data after pre-processing is sent into feature extraction network Backbone and image feature extraction is carried out, output multi-scale feature map, and it is input into integrated collection feature Neck module to strengthen spatial feature and semantic feature fusion, output enhanced feature map;Fusion enhanced feature map is sent into encoder, and global feature information is highlighted through the multi-scale receptive field of hollow convolution path, finally, the class and boundary box prediction of unmanned aerial vehicle target are generated after decoder generation.The present application is based on DINO architecture, and a high-precision unmanned aerial vehicle target detection model is constructed by a multi-scale hollow convolution fusion method.The detection accuracy and robustness of small-size unmanned aerial vehicle targets are improved by effectively extracting multi-scale features and recursively fusing.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, and specifically relates to a UAV target detection method based on feature fusion DINO. Background Technology

[0002] With the rapid development of drone technology, its applications in military reconnaissance, environmental monitoring, and traffic management are becoming increasingly widespread. Therefore, counter-drone technology has become a current research hotspot. When radar detects at low altitudes, radar waves are easily interfered with by ground reflections, and the curvature of the Earth limits the radar's detection range, making low-altitude targets easily obscured. Furthermore, numerous radio signals and electronic devices in cities interfere with radio frequency signals, making it difficult to accurately identify the radio frequency characteristics of drones. These problems result in insufficient reliability and stability of existing technologies in complex environments, making it difficult to meet the requirements of counter-drone missions.

[0003] Currently, radar, audio / video, and radio frequency technologies are widely used in UAV reconnaissance, with target detection technology playing a crucial role. Traditional target detection methods face numerous challenges when dealing with UAV targets. Firstly, UAV targets are small and vary greatly in scale, making them prone to false positives or false negatives in complex backgrounds. Secondly, traditional methods rely on manually designed features, resulting in poor classification and localization accuracy, failing to meet the demands for high-precision and robust detection. Furthermore, existing methods suffer from insufficient feature representation and imbalanced feature information flow, problems that become more pronounced in complex backgrounds and scenarios with significant scale variations. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems by providing a UAV target detection method based on feature fusion DINO, which aims to improve the accuracy and reliability of target detection by avoiding false detections and missed detections in existing target detection methods when detecting UAV targets.

[0005] The technical solution adopted in this invention is as follows: A UAV target detection method based on feature fusion DINO, the method comprising the following steps:

[0006] Image data preprocessing: acquire raw drone images and preprocess the raw drone image data;

[0007] Image data features are extracted by inputting the preprocessed data into the Backbone feature extraction network for image feature extraction and outputting multi-scale feature maps.

[0008] To reduce semantic feature gaps, feature maps are recursively fused. Multi-scale feature maps are input into the feature collection Neck module to enhance the fusion of spatial and semantic features, and an enhanced feature map is output.

[0009] The detection head calculates and outputs the detection results, flattens the fused features, and inputs them into the encoder in combination with the position encoding for feature enhancement. It highlights global feature information through dilated convolutional paths of multi-scale receptive fields, and inputs the enhanced feature map into the decoder for target localization and classification, and finally outputs the target detection results.

[0010] It is important to note that the Backbone is the backbone network used to extract features, while the Neck is an intermediate layer connecting the Backbone and the Head, used to reduce or adjust the dimensionality of features from the Backbone. Both are common terms in neural networks.

[0011] Furthermore, image data preprocessing includes data augmentation and data normalization.

[0012] Furthermore, to facilitate feature map fusion, the feature maps are divided into four layers. The Backbone residual network outputs a series of multi-scale feature maps, which are labeled as M1, M2, M3, and M4 in descending order of size.

[0013] Furthermore, the fusion of feature maps and semantic features specifically includes the following steps:

[0014] Feature map generation: Input feature maps M1, M2, M3, and M4 into the first column of the Neck module, and perform feature fusion through a Top-Down path to generate the first feature map; input the first feature map into the second column of the Neck module, and perform feature fusion through a Down-Top path to generate the second feature map;

[0015] The second feature map is recursively processed by the multi-scale dilated convolution fusion MACF module to obtain the third feature map. Then, the previous recursion is performed to obtain the fourth feature map.

[0016] Final feature map output: The fourth feature map is further fused through a Down-Top path to output the final feature map.

[0017] It is important to note that Multi-scale Atrous Convolution Fusion (MACF) is abbreviated as MACF.

[0018] Furthermore, the fused feature maps contain spatial information at different scales. To facilitate further feature processing in subsequent network layers, the Multi-Scale Dilated Convolutional Fusion (MACF) module captures contextual information of the input feature maps at multiple scales by performing dilated convolutions at different dilation rates. Feature maps are then generated through normalization and activation functions, and the multi-channel feature maps are fused. The specific steps include the following:

[0019] The input feature map is subjected to dilated convolution to capture contextual information at different scales;

[0020] Calculate the feature map size:

[0021] ;

[0022] in, The size of the output feature map. The size of the input feature map. For fill size, The void ratio, The kernel size is the convolution kernel size. This represents the convolution sliding stride;

[0023] And for , and Assignment:

[0024] ;

[0025] ;

[0026] in, The void ratio is multi-scale;

[0027] And set:

[0028] ;

[0029] ;

[0030] in, For the height of the output feature map, For the height of the input feature map, To determine the width of the output feature map, The width of the input feature map;

[0031] Each feature map is divided into four dilated convolutional layers, each connected to a batch normalization layer and a ReLU function to generate a third feature map. The third feature maps are merged by channel and fused with semantic features to obtain the final feature map. The final feature map contains spatial information at different scales.

[0032] Furthermore, in order to extract deeper and more abstract feature representations, reduce semantic gaps, and fuse high-level and low-level features, the current layer of the encoder receives the output of the feature map of the previous layer, and at the same time receives the feature map of the previous two layers after MACF processing.

[0033] Furthermore, the decoder receives the output of each layer of the feature map and the location information after the feature map is encoded and decoded, and outputs the detection result of the target.

[0034] Furthermore, the Neck module includes a recursive MACF structure to enhance the fusion of feature maps and semantic features.

[0035] Furthermore, the encoder incorporates an interlayer MACF structure to enhance feature representation.

[0036] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0037] 1. Based on the DINO architecture, this invention constructs a high-precision UAV target detection model through a multi-scale dilated convolution fusion method. By effectively extracting and recursively fusing multi-scale features, it reduces semantic gaps and improves the detection accuracy and robustness of small-sized UAV targets.

[0038] 2. This invention is based on dilated convolution technology, which achieves efficient feature fusion without significantly increasing computing resources, effectively reducing dependence on computing power and energy consumption, and providing feasible support for practical deployment. Attached Figure Description

[0039] Figure 1 This is a framework diagram of the present invention;

[0040] Figure 2 This is a diagram of the multi-scale dilated convolutional MACF structure of the present invention;

[0041] Figure 3 This is a diagram of the Transformer layer architecture of the present invention;

[0042] Figure 4 This is a schematic diagram of the drone detection results. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings.

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] This invention discloses a UAV target detection method based on DINO feature fusion, applicable to machine learning and electronic warfare fields. This method aims to optimize the detection of small-sized UAV targets with large scale variations. Based on the DINO architecture, it constructs a high-precision UAV target detection model through multi-scale dilated convolutional fusion technology. The specific steps are as follows: In the Backbone module, the input image is processed through a deep residual network to extract features, generating multi-scale spatial and semantic features. In the Neck module, the feature maps are fused layer by layer, and the feature representation is further enhanced through recursive paths. In the Encoder module, the feature maps are flattened into one-dimensional vectors and combined with position encoding for feature enhancement, and global feature information is highlighted through additional dilated convolutional paths with multiple receptive field sizes. In the Decoder module, the enhanced features are fed into the decoder, and the detection head outputs the position and category information of the UAV target. This invention effectively extracts and fuses multi-scale features, reduces semantic gaps, and improves the detection accuracy and robustness of small-sized UAV targets.

[0046] like Figure 1 As shown, a UAV target detection method based on feature fusion DINO is proposed, the method comprising the following steps:

[0047] Step S100: Acquire raw UAV image data and perform preprocessing;

[0048] Step S200: Input the image data from step S100 into the feature extraction network Backbone for image feature extraction and output a multi-scale feature map;

[0049] Step S300: Input the multi-scale feature map into the Neck module to enhance the fusion of spatial and semantic features, and output the enhanced feature map;

[0050] Step S400: Flatten the fused features and input them into the encoder in combination with the position encoding for feature enhancement. Highlight global feature information through the dilated convolution path of the multi-scale receptive field.

[0051] Step S500: The detection head calculates and outputs the detection results. The enhanced feature map is input into the decoder for target localization and classification. Finally, the detection results such as the target's position and category are output.

[0052] Specifically, step S300 includes the following steps:

[0053] Step S301: Generation of multi-scale feature maps: The input image is processed by the feature extraction network Backbone to obtain a series of multi-scale feature maps, which are labeled as M1, M2, M3 and M4 in descending order of size;

[0054] Step S302: Feature fusion using the top-down path: Input feature maps M1, M2, M3, and M4 into the first column of the Neck and perform feature fusion using the top-down path. The rule is: the initial feature map M4 is directly assigned to F4, and then through layer-by-layer fusion, F3=F4+M3, F2=F3+M2, F1=F2+M1, where "+" indicates that the corresponding element values ​​are added together.

[0055] Step S303: Feature fusion using the Down-Top path: The second column of the Neck uses the Down-Top path for feature fusion. The rule is: the initial feature map F1 is directly assigned to P1, and then through layer-by-layer fusion, P2=F2+P1, P3=F3+P2, P4=F4+P3;

[0056] Step S304: Feature enhancement of the recursive path: In the recursive path, P4 is first processed by multi-scale dilated convolution fusion with a MACF structure to obtain R4, and then the previous recursion is performed to obtain F3'=R4+F4+B3; similarly, P3 is fused by multi-scale dilated convolution to obtain R3, and then F2'=R3+F3+B2; again, P2 is processed to obtain F1'=R2+F2+B1;

[0057] Step S305: Output of final feature maps: F1', F2', and F3' are further fused through the above Down-Top path to finally output feature maps O1, O2, O3, and O4.

[0058] like Figure 2 As shown, the Multi-Scale Dilated Convolutional Fusion (MACF) module captures contextual information of the input feature map at multiple scales by performing dilated convolutions at different dilation rates. Then, it generates feature maps through batch normalization and activation functions, and fuses the multi-channel feature maps. Step S304 specifically includes the following steps:

[0059] Step S310: The input feature map first goes through a series of dilated convolutional layers to capture contextual information at different scales;

[0060] Step S311: Each dilated convolutional layer uses a 3x3 convolutional kernel with a multi-scale dilation rate (d).

[0061] Feature map size calculation:

[0062] ;

[0063] in, The size of the output feature map. The size of the input feature map. For fill size, The void ratio, The kernel size is the convolution kernel size. This represents the convolution sliding stride;

[0064] And for , and Assignment:

[0065] ;

[0066] ;

[0067] in, The void ratio is multi-scale;

[0068] And set:

[0069] ;

[0070] ;

[0071] in, For the height of the output feature map, For the height of the input feature map, To determine the width of the output feature map, The width of the input feature map;

[0072] therefore, This configuration can increase the receptive field of the convolutional layers while keeping the output feature map size constant, thereby improving model performance. The first layer uses a dilation rate... The standard convolution, the second layer uses dilation. The dilated convolution, the third layer uses dilation rate The dilated convolution, the fourth layer uses dilation rate Aperture convolution;

[0073] Step S312: Divide each feature map into four dilated convolutional layers, each layer is connected to a batch normalization layer and a ReLU function to generate a third feature map; the third feature map is merged by channel and fused with semantic features to obtain the final feature map; the final feature map contains spatial information at different scales.

[0074] like Figure 3 As shown, in step S500, the current layer of the encoder receives the output of the previous layer M(j-1), and also receives the feature map of M(j-2) after MACF processing; it extracts deeper and more abstract feature representations, and fuses the features of the high-level and low-level layers to reduce semantic gap. Figure 3 In The notation indicates a multi-layered architecture—the output of each layer becomes the input of the next layer. The original input text is transformed into an input embedding through an embedding layer. This embedding layer finds a corresponding vector for each token. After obtaining the token embedding, positional encoding is added. This positional encoding helps the model understand the relative position of the token.

[0075] Simultaneously utilizing both low-level and high-level features, predictions are performed at different layers. This is because an image may contain multiple targets of varying sizes, and different features may be needed to distinguish different targets. Simple targets can be detected using only shallow features, while complex targets require more complex features. The entire process first performs depthwise convolution on the original image, and then performs predictions at different feature layers. Its advantage is that it outputs the corresponding target at different layers, eliminating the need to go through all layers (i.e., for some targets, unnecessary forward operations are not required), which can accelerate the network to some extent and improve the algorithm's detection performance. Its disadvantage is that the obtained features are not robust, consisting of weak features because many features are obtained from shallower layers.

[0076] Finally, the decoder outputs the target's localization and classification information, ultimately outputting the target detection result, such as... Figure 4 As shown, Figure 4 The left side shows the detection results without using the detection method of this invention. Figure 4 The right side shows the detection results using the detection method of this invention. Figure 4 It can be seen that the UAV target detection method based on feature fusion DINO can accurately capture UAV targets and reduce false detection and false detection rates.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1.A UAV target detection method based on feature fusion DINO, characterized in that, The method comprises the following steps: The image data preprocessing, obtaining the original unmanned aerial vehicle image, and preprocessing the original unmanned aerial vehicle image data; Extracting image data features, inputting the preprocessed data into the feature extraction network Backbone to extract image features, and outputting multi-scale feature maps; Feature map recursive fusion, inputting the multi-scale feature map into the integration and collection feature Neck module to strengthen the fusion of spatial features and semantic features, and outputting an enhanced feature map; The detection head calculates and outputs the detection result, flattens the fused features, and inputs the position encoding into the encoder for feature enhancement, highlights the global feature information through the multi-scale receptive field of the empty convolution path, and inputs the enhanced feature map into the decoder for target positioning and classification, and finally outputs the detection result of the target; The image data preprocessing includes data enhancement and normalization operation on the data; The Backbone residual network outputs a series of multi-scale feature maps, which are sequentially labeled as M1, M2, M3, and M4 according to the size from large to small; The feature map recursive fusion specifically comprises the following steps: Generating a feature map: inputting the M1, M2, M3, and M4 feature maps into the first column of the Neck module, performing feature fusion through the Top-Down path to generate a first feature map; inputting the first feature map into the second column of the Neck module, performing feature fusion through the Down-Top path to generate a second feature map; Recursively processing the second feature map through the multi-scale empty convolution fusion MACF module to obtain a third feature map, and then performing a front recursive operation to obtain a fourth feature map; Finally, the fourth feature map is further fused through the Down-Top path to output the final feature map. 2.The UAV target detection method based on feature fusion DINO of claim 1, wherein, The multi-scale empty convolution fusion MACF module captures the context information of the input feature map at multiple scales by performing empty convolution at different hole rates, generates a feature map through normalization and activation function, and fuses the multi-channel feature map, and the specific steps include the following: Input feature map is processed by empty convolution to capture context information at different scales; Calculate the size of the feature map: ； wherein, is the size of the output feature map, is the size of the input feature map, is the padding size, is the hole rate, is the size of the convolution kernel, is the convolution sliding step. and for , and assignment: ； ； wherein, is the void fraction on a multiscale; And set: ； ； wherein, is the height of the output feature map, is the height of the input feature map, is the width of the output feature map, is the width of the input feature map; Each feature map is divided into four layers of empty convolution layers, each layer is connected with a batch normalization layer and a ReLU function, which is used to generate a third feature map; the third feature map is merged by channel, fused with semantic features, and obtains a final feature map; the final feature map contains spatial information at different scales. 3.The UAV target detection method based on feature fusion DINO according to claim 2, characterized in that, The current layer of the encoder receives the output of the previous layer feature map, and simultaneously receives the feature map processed by the MACF of the previous two layers. 4.The method of claim 3, wherein, The decoder receives the output of each layer of the feature map and the position positioning information after decoding the position encoding of the feature map, and outputs the detection result of the target. 5.The UAV target detection method based on feature fusion DINO of claim 1, wherein, The Neck module is provided with a recursive MACF structure for strengthening the fusion of feature maps and semantic features. 6.The UAV target detection method based on feature fusion DINO according to claim 3, characterized in that, The encoder is provided with an interlayer MACF structure for strengthening feature expression.

Images