Airport scene small target detection method and system based on adaptive distortion feature matching
By introducing an adaptive distortion feature matching detection method in the YOLO algorithm, using the deformable and attention feature extraction module, combined with dual-path aggregation and NWD loss function, the viewing angle distortion and small target recognition problems in airport scene object detection are solved. Accurate detection of aircraft, vehicles and personnel and high-precision identification of small targets are achieved.
Patent Information
- Application Number
- CN202411221739.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-11-22
AI Technical Summary
The existing YOLO algorithm faces the problems of viewing angle distortion, occlusion and insufficient recognition capabilities in airport scene object detection, resulting in poor detection accuracy and effectiveness.
Adopting the small object detection method for airport scenes with adaptive distortion feature matching, the object detection model including deformable feature extraction module and attention feature extraction module is introduced by combining the dual-path aggregation module and the detection head network. Function to improve detection accuracy.
Accurate detection of aircraft, vehicles and personnel in complex airport environments, adapt to perspective distortions and improve the accuracy of small target detection, enhancing the overall detection effect.
Smart Images

Figure CN119006796A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision technology, and in particular to an airport scene small target detection method and system with adaptive distortion feature matching. Background Art
[0002] Target detection refers to the detection and classification of targets of interest from images. Airport scene target detection refers to the detection and identification of different types of target objects in the airport, such as aircraft, vehicles, personnel, etc., by analyzing the image data of the airport scene. The purpose of airport scene target detection is to obtain real-time information of the airport scene to support the operation, management and service of the airport. By detecting and identifying targets on the airport scene, we can understand the traffic conditions, passenger flow distribution, and the use of equipment and facilities in the airport, thereby providing data support for the optimization and improvement of the airport. This type of problem has the characteristics of complex background, perspective distortion, target occlusion, and small target scale. Accurately detecting and identifying targets in different scenes is of great significance to airport management.
[0003] YOLO is a commonly used target detection method. The algorithm achieves efficient target detection by dividing the input image into multiple grids and performing target classification prediction and bounding box regression in a single network. However, the existing YOLO algorithm faces multiple challenges in airport scene target detection. First, there is perspective distortion in large scenes, and YOLO performs poorly when identifying targets of irregular shapes. Secondly, there is often occlusion in the airport environment, which causes the target to be partially occluded, affecting the accuracy of detection. Finally, YOLO's recognition ability for small objects is relatively weak, which can easily lead to missed detection or false detection, limiting the overall detection effect. How to improve the accuracy of the YOLO algorithm in complex background and small target detection tasks has become an important development direction for airport scene target detection. Summary of the invention
[0004] The present invention aims at solving the problems in the prior art and provides an airport scene small target detection method and system with adaptive distortion feature matching.
[0005] The technical solution adopted by the present invention is: a method for detecting small targets at airports by adaptive distortion feature matching, comprising the following steps:
[0006] Step 1: Obtain scene image data to form a data set, annotate the data set and perform preprocessing;
[0007] Step 2: Build a target detection model, which includes a backbone network and a detection head network;
[0008] The backbone network includes a deformable feature extraction module and an attention feature extraction module for feature extraction;
[0009] The detection head network includes a dual-path aggregation module for fusing feature maps extracted by the backbone network and a detection head for detecting the fused feature maps;
[0010] Step 3: Train the target detection model in step 2 to obtain a trained target detection model;
[0011] Step 4: Use the trained object detection model to detect the scene image data.
[0012] Further, the deformable feature extraction module includes N deformable convolution blocks and an efficient layer aggregation network ELAN connected in sequence;
[0013] The deformable convolution block consists of a deformable convolution layer, a normalization layer, and an activation layer.
[0014] Further, the attention feature extraction module includes an attention feature extractor AttFE and a P-AttFE module connected in sequence;
[0015] AttFE includes the maximum pooling network, ELAN, and the compressed activation network SENet;
[0016] The P-AttFE module includes a maximum pooling network, ELAN, SPPCSPC modules, and a compressed activation network SENet.
[0017] Furthermore, the two AttFE and P-AttFE modules both obtain feature maps, which are the first feature map, the second feature map and the third feature map, respectively.
[0018] Furthermore, the dual-path aggregation module includes dual paths, the first path is for fusing from a low-resolution feature map to a high-resolution feature map; the second path is for fusing from a high-resolution feature map to a low-resolution feature map.
[0019] Furthermore, the first path is to fuse the third feature map with the second feature map to obtain a fourth feature map, and fuse the fourth feature map with the first feature map to obtain a fifth feature map;
[0020] The second path is to fuse the fifth feature map with the fourth feature map to obtain the sixth feature map, and to fuse the sixth feature map with the third feature map to obtain the seventh feature map.
[0021] Furthermore, the detection head includes a feature representation module and a convolutional layer.
[0022] Furthermore, the loss function used in the training process is a joint loss function, and the calculation method of the joint loss function is as follows:
[0023] L=W loc ×L loc +W conf×L conf +W cls ×L cls
[0024] Where: L is the joint loss function, L loc is the coordinate regression loss, L conf is the confidence loss, L cls is the classification loss; W loc , W conf , W cls They are the proportion of coordinate regression loss, confidence loss and classification loss in the total loss respectively;
[0025] in,
[0026] L loc =W CIoU ×(1-CIoU)+W NWD ×(1-NWD)
[0027] Where: NWD is the NWD loss function, CIoU is the CIoU loss function, W CIoU and W NWD is the corresponding weight.
[0028] Furthermore, the scene image data obtained in step 1 constitutes a data set, and the data set is annotated and preprocessed as follows:
[0029] Get the collected video, extract images at intervals of n frames, and select the required images to form a data set;
[0030] Labeling of objects on the airport scene as aircraft, vehicles, and personnel;
[0031] The preprocessing includes:
[0032] The images in the dataset are enhanced by random perspective transformation, color space enhancement, and random occlusion.
[0033] A detection system for a small target detection method at an airport scene based on adaptive distortion feature matching, comprising:
[0034] Input module for obtaining datasets;
[0035] A feature extraction module for performing feature extraction; the feature extraction module includes a deformable feature extraction module and an attention feature extraction module;
[0036] A dual-path aggregation module for fusing feature maps output by the feature extraction module;
[0037] A detection module used to detect the fused feature map.
[0038] The beneficial effects of the present invention are:
[0039] (1) The method of the present invention can adapt to the complex airport environment and realize accurate detection of aircraft, vehicles and personnel in the airport scene;
[0040] (2) The method of the present invention includes a deformable feature extraction module, introduces deformable convolution, and adaptively learns deformable features, which solves the perspective distortion problem of surveillance cameras in large airport scenes;
[0041] (3) The method of the present invention includes an attention feature extraction module, introduces a channel attention mechanism, and retains task-oriented features and suppresses interference features by adaptively allocating importance weights of different features;
[0042] (4) The loss function provided by the present invention introduces NWD loss into the coordinate regression loss, and improves the detection accuracy of small targets in the airport scene by measuring the similarity of the Gaussian distribution of the predicted box and the true box. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 The figure is a flow chart of the method of the present invention.
[0044] Figure 2 It is a structural schematic diagram of the target detection model in the present invention. DETAILED DESCRIPTION
[0045] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0046] like Figure 1 As shown, a method for detecting small targets at airports based on adaptive distortion feature matching includes the following steps:
[0047] Step 1: Obtain scene image data to form a data set, annotate the data set and perform preprocessing;
[0048] Get the collected video, extract images at intervals of n frames, and select the required images to form a data set;
[0049] Labeling of objects on the airport scene as aircraft, vehicles, and personnel;
[0050] The preprocessing includes:
[0051] The images in the dataset are enhanced by random perspective transformation, color space enhancement, and random occlusion.
[0052] Step 2: Build a target detection model, which includes a backbone network and a detection head network;
[0053] The backbone network includes a deformable feature extraction module and an attention feature extraction module for feature extraction;
[0054] The deformable feature extraction module includes N deformable convolution blocks connected in sequence and an efficient layer aggregation network ELAN;
[0055] The deformable convolution block consists of a deformable convolution layer, a normalization layer, and an activation layer.
[0056] The attention feature extraction module includes an attention feature extractor AttFE and a P-AttFE module connected in sequence;
[0057] AttFE includes the maximum pooling network, ELAN, and the compressed activation network SENet;
[0058] The P-AttFE module includes a maximum pooling network, ELAN, SPPCSPC modules, and a compressed activation network SENet.
[0059] Both AttFE and P-AttFE modules obtain feature maps, namely the first feature map, the second feature map and the third feature map.
[0060] The detection head network includes a dual-path aggregation module for fusing feature maps extracted by the backbone network and a detection head for detecting the fused feature maps;
[0061] The dual-path aggregation module includes dual paths, the first path is to fuse from a low-resolution feature map to a high-resolution feature map; the second path is to fuse from a high-resolution feature map to a low-resolution feature map.
[0062] The first path is to fuse the third feature map with the second feature map to obtain the fourth feature map, and fuse the fourth feature map with the first feature map to obtain the fifth feature map;
[0063] The second path is to fuse the fifth feature map with the fourth feature map to obtain the sixth feature map, and to fuse the sixth feature map with the third feature map to obtain the seventh feature map.
[0064] The detection head consists of a feature representation module and a convolutional layer.
[0065] Step 3: Train the target detection model in step 2 to obtain a trained target detection model;
[0066] The loss function used in the training process is the joint loss function, and the calculation method of the joint loss function is as follows:
[0067] L=W loc ×L loc +W conf ×L conf +W cls ×L cls
[0068] Where: L is the joint loss function, L loc is the coordinate regression loss, L conf is the confidence loss, L cls is the classification loss; W loc , W conf , W cls They are the proportion of coordinate regression loss, confidence loss and classification loss in the total loss respectively;
[0069] in,
[0070] L loc =W CIoU ×(1-CIoU)+W NWD ×(1-NWD)
[0071] Where: NWD is the NWD loss function, CIoU is the CIoU loss function, W CIoU and W NWD is the corresponding weight.
[0072] Step 4: Use the trained object detection model to detect the scene image data.
[0073] A detection system for a small target detection method at an airport scene based on adaptive distortion feature matching, comprising:
[0074] Input module for obtaining datasets;
[0075] A feature extraction module for performing feature extraction; the feature extraction module includes a deformable feature extraction module and an attention feature extraction module;
[0076] A dual-path aggregation module for fusing feature maps output by the feature extraction module;
[0077] A detection module used to detect the fused feature map.
[0078] The present invention will be further described below in conjunction with specific embodiments.
[0079] A method for detecting small targets at airports based on adaptive distortion feature matching comprises the following steps:
[0080] Step 1: Obtain scene image data to form a data set, annotate the data set and perform preprocessing;
[0081] First, the video data collected by the airport scene surveillance camera is collected;
[0082] The videos are collected from 9 surveillance cameras at 4 airports, and the videos of each airport cover all locations at the airport.
[0083] Images of various scenes are selected as data and annotated. The images to be detected are preprocessed and divided into training set, verification set and test set.
[0084] 1) Read the collected video frame by frame and extract images at intervals of 5 frames;
[0085] 2) A total of 5064 images with different forms of scene objects, different lighting conditions, camera perspectives and weather conditions are selected from the extracted images to form a data set; the morphological diversity includes different types of aircraft, different types of vehicles (cars, buses, etc.), and different groups of people; the lighting conditions cover daytime, dusk, night, etc.; the camera perspectives are different shooting angles, such as front, side, and top views; the weather conditions include sunny, rainy, foggy, and other meteorological conditions.
[0086] 3) The objects in the airport scene are labeled into three categories: aircraft, vehicles and personnel. The annotation information includes: the category of each target, the percentage of the x-coordinate of the center point of the detection frame in the width of the whole picture, the percentage of the y-coordinate in the height of the whole picture, the percentage of the detection frame width in the width of the whole picture, and the percentage of the detection frame height in the height of the whole picture. The dataset finally contains 13,127 target aircraft, 8,663 target vehicles and 8,946 target personnel.
[0087] 4) Convert all image resolutions to 640×640, and use random perspective transformation, color space enhancement, and random occlusion to enhance the images to be detected and the corresponding annotated images. Finally, divide the dataset into training set, validation set, and test set according to 8:1:1.
[0088] Step 2: Build a target detection model. The target detection model includes a backbone network and a detection head network, such as Figure 2 As shown;
[0089] The backbone network includes a deformable feature extraction module and an attention feature extraction module for feature extraction;
[0090] The deformable feature extraction module includes four deformable convolutional blocks (DBS) connected in sequence and an efficient layer aggregation network ELAN (Efficient Layer Aggregation Network); the DBS block consists of a deformable convolution DCNv2 (DeformableConvolutional Network), batch normalization (BN—BatchNorm) and SiLU (Sigmoid Linear Unit) activation function. DBS is based on the CBS block in the original YOLOv7, on which the standard convolution is replaced by a deformable convolution DCNv2.
[0091] First, the original image with a dimension of 640×640×3 is input into the first DBS, and then each DBS extracts and transforms the input data. The second and fourth DBS downsample the input image, and the number of output channels of the four DBSs are {32, 64, 64, 128}, so the dimension of the feature map after DBS processing is 160×160×128. Through DBS processing, the original data is converted into information-rich deformable features, providing a sufficient feature basis for subsequent long-distance information aggregation. Next, the output of the last DBS is sent to the ELAN module. The ELAN module keeps the input and output resolution unchanged and doubles the number of channels, so the output dimension of the deformable feature extraction module is 160×160×256.
[0092] The attention feature extraction module is composed of two attention feature extractors (AttFE—Attention Feature Extractor) and an improved P-AttFE in series. Each AttFE consists of a multi-layer structure, including a maximum pooling network (MP1), ELAN, and a squeeze activation network (SENet—Squeeze-and-Excitation Network). MP1 aims to extract more image features in the process of downsampling the feature map. The role of SENet is to improve the expression ability of important features by adaptively adjusting the weights of feature channels. Based on AttFE, P-AttFE introduces the SPPCSPC module between the ELAN and SE modules. This module combines the advantages of spatial pyramid pooling (SPP—Spatial Pyramid Pooling) and cross-stage local network (CSPN—Cross Stage Partial Network) to improve the multi-scale feature expression ability and computational efficiency of the model.
[0093] The cascaded attention feature extraction module takes the output of the deformable feature extraction module as input, and obtains the first feature map, the second feature map, and the third feature map respectively after two AttFEs and one P-AttFE. AttFE downsamples the feature map, and the number of output channels is twice the number of input channels. P-AttFE downsamples the feature map, and the number of output channels is half the number of input channels. Therefore, the dimensions of the first feature map, the second feature map, and the third feature map are 80×80×512, 40×40×1024, and 20×20×512, respectively.
[0094] The detection head network includes a dual-path aggregation module for fusing feature maps extracted by the backbone network and a detection head for detecting the fused feature maps;
[0095] The dual-path aggregation module realizes bidirectional fusion aggregation of features through two reverse paths. The first path is the fusion path from low-resolution feature maps to high-resolution feature maps; the second path is the fusion path from high-resolution feature maps to low-resolution feature maps.
[0096] The purpose of the first path is to integrate the semantic information in the low-resolution feature map into the high-resolution feature map to enhance the semantic expression ability of the high-resolution feature map. It includes ELAN-H and the upsampling module Up. ELAN-H has the same network structure as ELAN, but the difference is that it fuses information from different layers. The upsampling module Up is responsible for upsampling the low-resolution feature map to the size of the high-resolution feature map to facilitate feature fusion between the two.
[0097] The third feature map first passes through the Up module to double its resolution and halve the number of channels. The second feature map is processed by the CBS block with the same resolution and the number of channels reduced to one-fourth of the original. Next, the third feature map with enlarged resolution is added element by element to the second feature map processed by the CBS block, and the result is sent to ELAN-H for feature fusion and enhancement. After processing by ELAN-H, a new set of feature maps, namely the fourth feature map, is obtained. The ELAN-H module keeps the input and output resolution and the number of channels unchanged, so the dimension of the fourth feature map is 40×40×256. After the fourth feature map and the first feature map undergo the same fusion operation, the fifth feature map is obtained, and the dimension of the fifth feature map is 80×80×128. At this point, the fusion from low-resolution feature maps to high-resolution feature maps is completed.
[0098] The purpose of the second path is to fuse the detail information in the high-resolution feature map into the low-resolution feature map to enhance the detail expression ability of the low-resolution feature map. This path also contains two key modules: ELAN-H module and MP2 module. The difference between MP2 and MP1 is that the number of output channels of the former is twice that of the input, while the number of input and output channels of MP1 remains unchanged.
[0099] The fifth feature map is input into the MP2 module for downsampling to match the resolution of the fourth feature map. The input and output resolution of the MP2 module is halved and the number of channels is doubled. Next, the downsampled feature map is added element by element to the fourth feature map, and the result is sent to ELAN-H for effective feature aggregation. The output is the sixth feature map with a dimension of 40×40×256. The sixth feature map and the third feature map undergo the same fusion operation to obtain the seventh feature map with a dimension of 20×20×512, completing dual-path fusion.
[0100] The detection head consists of three, and the fifth feature map, the sixth feature map, and the seventh feature map are respectively sent to the detection head adapted to different scales. Each detection head processes feature information from different levels and resolutions through a feature representation (REP-Representation) module and a convolutional layer combination to effectively represent features of different scales. Finally, three sets of images with different resolutions are output to obtain candidate boxes of detection objects of different scales on the airport scene and the categories of targets in the candidate boxes. The dimensions of the output images of the detection heads corresponding to the fifth feature map, the sixth feature map, and the seventh feature map are 80×80×(3+5)×3, 40×40×(3+5)×3, and 20×20×(3+5)×3, respectively, to predict small, medium, and large targets in the image, respectively.
[0101] Step 3: Train the target detection model in step 2 to obtain a trained target detection model;
[0102] The present invention introduces the NWD (Normalized Gaussian Wasserstein Distance) loss function to enhance the detection capability of small targets.
[0103] The joint loss function consists of three parts: coordinate regression loss, confidence loss, and classification loss, as shown in the following formula.
[0104] The loss function is a joint loss function, and the calculation method of the joint loss function is as follows:
[0105] L=W loc ×L loc +W conf ×L conf +W cls ×L cls
[0106] Where: L is the joint loss function, L loc is the coordinate regression loss, L conf is the confidence loss, L cls is the classification loss; W loc , W conf , W cls are the proportion of coordinate regression loss, confidence loss and classification loss in the total loss respectively; in this embodiment, they are 0.05, 0.1 and 0.125 respectively.
[0107] Among them, the coordinate regression loss is used to calculate the difference between the predicted bounding box and the true bounding box, using the CIoU (Complete Intersection over Union) loss function and the NWD loss function for calculation. The calculation method is shown in the following formula.
[0108] L loc=W CIoU ×(1-CIoU)+W NWD ×(1-NWD)
[0109] Where: NWD is the NWD loss function, CIoU is the CIoU loss function, W CIoU and W NWD are the corresponding weights, which are 0.5 and 0.5 respectively in this embodiment.
[0110] Confidence loss is used to measure the confidence of the predicted bounding box and is calculated using the BCE (Binary Cross-Entropy) loss function. Classification loss is used to calculate the loss between the predicted category and the true category and is calculated using the BCE loss function.
[0111] Traditional CIoU shows certain limitations when dealing with objects of different scales. For larger objects, small position deviations have less impact on IoU measurements. In contrast, for small objects, the same position deviation will cause the IoU value to drop significantly, affecting the accurate assignment of target labels. Overall, CIoU can provide more accurate positioning evaluation for larger objects, but it does not meet the expected performance when evaluating the position accuracy of smaller objects, thus affecting the model's detection accuracy for small objects.
[0112] The NWD loss function is introduced into the coordinate regression loss to improve the detection accuracy of small objects. The NWD loss is calculated in the following two steps: first, the predicted box and the true box are modeled as a two-dimensional Gaussian distribution, and then the Wasserstein distance is applied to measure the similarity of the Gaussian distribution of the predicted box and the true box. As shown below:
[0113]
[0114] Where: X p and Y p Respectively represent the horizontal and vertical coordinates of the upper left corner of the detection box; X g and Y g Respectively represent the horizontal and vertical coordinates of the upper left corner of the real box; W p and H p Respectively represent the width and height of the prediction box; W g and H g Represent the width and height of the real box respectively.
[0115] The model is trained for 150 rounds and the weights of the optimal model are saved.
[0116] Step 4: Use the trained object detection model to detect the scene image data.
[0117] The target detection model constructed by the present invention is obtained by improving the existing YOLOv7 model, which can achieve high-precision detection of airport scene targets in complex airports, and is of great significance to airport operations and management. The present invention can achieve accurate detection of aircraft, vehicles and personnel in airport scenes. Deformable convolution is introduced in the deformable feature extraction module to adaptively learn deformable features, effectively solving the perspective distortion problem of surveillance cameras in large airport scenes. The channel attention mechanism is introduced in the attention feature extraction module to adaptively assign importance weights of different features, retain task-oriented features, and suppress interference features. The NWD loss is introduced into the coordinate regression loss, and the detection accuracy of small targets in the airport scene is improved by measuring the similarity of the Gaussian distribution of the predicted box and the true box.
Claims
1. A method for detecting small targets at airports based on adaptive distortion feature matching, characterized in that: The following steps are involved: Step 1: Obtain scene image data to form a data set, annotate the data set and perform preprocessing; Step 2: Build a target detection model, which includes a backbone network and a detection head network; The backbone network includes a deformable feature extraction module and an attention feature extraction module for feature extraction; The detection head network includes a dual-path aggregation module for fusing feature maps extracted by the backbone network and a detection head for detecting the fused feature maps; Step 3: Train the target detection model in step 2 to obtain a trained target detection model; Step 4: Use the trained object detection model to detect the scene image data.
2. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 1, characterized in that: The deformable feature extraction module includes N deformable convolution blocks and an efficient layer aggregation network ELAN connected in sequence; The deformable convolution block consists of a deformable convolution layer, a normalization layer, and an activation layer.
3. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 2, characterized in that: The attention feature extraction module includes an attention feature extractor AttFE and a P-AttFE module connected in sequence; AttFE includes the maximum pooling network, ELAN, and the compressed activation network SENet; The P-AttFE module includes a maximum pooling network, ELAN, SPPCSPC modules, and a compressed activation network SENet.
4. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 3, characterized in that: The two AttFE and P-AttFE modules both obtain feature maps, which are the first feature map, the second feature map and the third feature map respectively.
5. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 4, characterized in that: The dual-path aggregation module includes dual paths, a first path for fusing from a low-resolution feature map to a high-resolution feature map, and a second path for fusing from a high-resolution feature map to a low-resolution feature map.
6. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 5, characterized in that: The first path is to fuse the third feature map with the second feature map to obtain the fourth feature map, and fuse the fourth feature map with the first feature map to obtain the fifth feature map; The second path is to fuse the fifth feature map with the fourth feature map to obtain the sixth feature map, and to fuse the sixth feature map with the third feature map to obtain the seventh feature map.
7. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 6, characterized in that: The detection head includes a feature representation module and a convolutional layer.
8. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 1, characterized in that: The loss function used in the training process is a joint loss function, and the calculation method of the joint loss function is as follows: L=W loc ×L loc +W conf ×L conf +W cls ×L cls Where: L is the joint loss function, L loc is the coordinate regression loss, L conf is the confidence loss, L cls is the classification loss; W loc , W conf , W cls They are the proportion of coordinate regression loss, confidence loss and classification loss in the total loss respectively; in, L loc =W CIoU ×(1-ClOU)+W NWD ×(1-NWD) Where: NWD is the NWD loss function, CIoU is the CIoU loss function, W CIoU and W NWD is the corresponding weight.
9. The method for detecting small targets at airports based on adaptive distortion feature matching according to claim 1, characterized in that: The scene image data obtained in step 1 constitutes a data set, and the data set is annotated and preprocessed as follows: Get the collected video, extract images at intervals of n frames, and select the required images to form a data set; Labeling of objects on the airport scene as aircraft, vehicles, and personnel; The pre-processing includes: The images in the dataset are enhanced by random perspective transformation, color space enhancement, and random occlusion.
10. A detection system using the airport scene small target detection method using adaptive distortion feature matching as claimed in any one of claims 1 to 9, characterized in that: include: Input module for obtaining datasets; A feature extraction module for performing feature extraction; The feature extraction module includes a deformable feature extraction module and an attention feature extraction module; A dual-path aggregation module for fusing feature maps output by the feature extraction module; A detection module used to detect the fused feature map.
Citation Information
Patent Citations
Lightweight airport scene monitoring method based on YOLOv4
CN115457465A
Semantic input sampling for explanation (SISE) of convolutional neural networks
US20220058431A1
Cited By
Train pantograph catenary arc discharge identification method, device and system and storage medium
CN119888605A
Method and device for detecting small target in complex environment based on unmanned aerial vehicle image
CN120032212A
