Airplane flight recorder detection and positioning method based on image recognition deep learning
Patent Information
- Application Number
- CN202611211730.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-10-16
AI Technical Summary
在复杂坠毁现场中，飞机飞行记录器可能散落在残骸、泥土、植被、水面、烧蚀痕迹或金属构件之间，存在目标尺度小、外观受损、局部遮挡、背景干扰强、拍摄角度多变、光照不均和目标数量少等问题
（1）本发明目标检测定位模型利用SPD-Conv模块通过空间到深度重排和非跨步卷积将空间信息转移至通道维度进行下采样，能够充分保留飞行记录器目标的边缘、纹理及局部颜色等早期细粒度特征，避免关键细节信息在下采样过程中丢失，从而改善了对受损、部分遮挡或外观变形飞行记录器目标的识别效果；同时将LSKA注意力机制融入SPPF模块得到SPPF-LSKA注意力模块，在多尺度池化获得的不同感受野上下文特征基础上，通过大核可分离卷积生成空间与通道双重注意力权重，对飞行记录器所在的关键区域进行特征增强，有效提升了模型在复杂背景干扰、局部遮挡和目标外观受损条件下的飞行记录器目标检测稳定性，还增设有小目标检测层，能够融合骨干网络浅层的细节特征，增强了对复杂坠毁现场图像中占比小的飞机飞行记录器目标感知能力，有效降低了漏检概率。
Smart Images
Figure CN122929443A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of civil aviation safety accident investigation, and in particular to a method for detecting and locating aircraft flight recorders based on image recognition deep learning. Background Technology
[0002] Aircraft flight recorders typically consist of a flight data recorder and a cockpit voice recorder, commonly known as a black box. They record flight parameters, crew voice recordings, and related operational information, serving as crucial evidence for aviation accident investigations, accident cause analysis, and safety improvements. After an accident, quickly locating the flight recorder can significantly improve search efficiency and buy time for subsequent investigations. In complex crash sites, flight recorders may be scattered among wreckage, soil, vegetation, water, ablation marks, or metal components, presenting challenges such as small target size, damaged appearance, partial obstruction, strong background interference, variable shooting angles, uneven lighting, and a limited number of targets. Traditional manual visual searches rely on human experience, resulting in low search efficiency and difficulty maintaining consistent accuracy in tasks involving drone inspections, rescue robot inspections, or large-scale image screening. Existing deep learning object detection algorithms can achieve end-to-end object recognition and localization. Among them, the YOLO series algorithms have advantages such as fast inference speed and convenient deployment. The original YOLO series algorithms extract semantic features through continuous downsampling in the backbone network. Details such as texture, edge, and color of small-sized targets are easily lost in the early downsampling process. At the same time, its detection head is mainly aimed at conventional-scale targets, and it is easy to miss flight recorder targets, which occupy a small proportion in large-format scene images. Therefore, it is necessary to provide an aircraft flight recorder target detection and localization model and method for searching for aircraft flight recorders in complex crash sites, which can improve the detection accuracy and recall rate of small targets, occluded targets, and aircraft flight recorder targets in complex backgrounds while maintaining lightweight and real-time performance. Summary of the Invention
[0003] The purpose of this invention is to provide a method for detecting and locating aircraft flight recorders based on deep learning for image recognition. This method can fully preserve early fine-grained features such as the edges, textures, and local colors of the flight recorder target, and avoid the loss of key details during downsampling. This improves the recognition effect on damaged, partially occluded, or deformed flight recorder targets, enhances the perception ability of small-sized aircraft flight recorder targets in large-format complex crash scene images, and effectively reduces the probability of missed detection.
[0004] The objective of this invention is achieved through the following technical solution: A method for detecting and locating aircraft flight recorders based on image recognition deep learning, the method comprising: S1. Obtain the search image sequence dataset, which includes image data and the location information corresponding to the images; construct a target detection and localization model with the improved YOLOv8n network architecture as the core, which includes a backbone network, a neck feature fusion network and a detection head, and the target detection and localization model is trained using the aircraft flight recorder image sample dataset for target detection and recognition. S2. Input the image data of the search image sequence dataset into the target detection and localization model for recognition. The backbone network performs multi-scale feature extraction on the image data layer by layer, and then performs attention weighting processing through the SPPF-LSKA attention module to output the weighted features. S3. The neck feature fusion network performs layer-by-layer feature fusion on the weighted features and the backbone network output features through the upsampling branch, and then performs layer-by-layer feature fusion through the downsampling branch to obtain several levels of fused features. S4. The detection head uses several levels of fusion features to perform target detection and recognition of the aircraft flight recorder and outputs the target detection results.
[0005] To better implement this invention, the target detection result includes the presence or absence of an aircraft flight recorder target and a target detection box, and also includes the following method: S5. Filter the image data in the image sequence dataset to find the target of the aircraft flight recorder. Store the image data and the corresponding target detection results. Extract the location information corresponding to the image data and parse the geographical location data of the aircraft flight recorder target.
[0006] Preferably, the backbone network of the target detection and localization model includes a four-layer structure, the upsampling branch and the downsampling branch of the neck feature fusion network each include a three-layer structure, and the detection head includes a small target detection head specifically adapted for target detection in aircraft flight recorders.
[0007] Preferably, the backbone network of the target detection and localization model has two SPD-Conv modules connected sequentially at the front end of the four-layer structure. The SPD-Conv module includes a spatial-to-depth transformation layer and a non-staggered convolutional layer with a stride of 1. The spatial-to-depth transformation layer performs interval sampling and rearrangement of the input features according to a preset scale factor s to obtain rearranged features. The non-staggered convolutional layer performs channel mapping on the rearranged features and outputs features. In the four-layer structure of the backbone network, the first three layers sequentially output features A1 to A3 at three scales. The last layer of the four-layer structure of the backbone network contains an SPPF-LSKA attention module. The SPPF-LSKA attention module extracts global multi-scale contextual information, calculates spatial and channel dual attention weights, and performs attention weight weighting on the output features of the last layer in the four-layer structure of the backbone network to output the weighted feature A4.
[0008] Preferably, the three-layer structure of the upsampling branch of the neck feature fusion network corresponds one-to-one with the first three layers of the four-layer structure of the backbone network. The three-layer structure of the upsampling branch is upsampled sequentially and spliced and fused with the corresponding layer of the backbone network to output three-layer scale features of features B2 to B4. The output feature B4 of the upsampling branch is input into the small target detection head of the detection head.
[0009] Preferably, the three-layer structure of the downsampling branch corresponds one-to-one with the three-layer structure of the upsampling branch. The three-layer structure of the downsampling branch performs downsampling processing in sequence and splices and fuses with the corresponding layer of the upsampling branch to output features C2 to C4. The detection head also includes three scale detection heads. The three scale detection heads receive features C2 to C4 in sequence and perform target detection and recognition processing of the aircraft flight recorder that adapts to scale changes.
[0010] Preferably, the detection head adopts a decoupled detection structure and performs target category detection and target detection box regression processing. The target category detection is whether it is an aircraft flight recorder target; if the target category is detected as an aircraft flight recorder target, the target detection box regression processing predicts the target detection box.
[0011] Preferably, the aircraft flight recorder image sample dataset includes aircraft flight recorder image samples, target category labels, and target bounding boxes. The aircraft flight recorder image samples contain target samples and empty target samples. Target samples are image samples with an aircraft flight recorder target, and the target category label of target samples has a record attribute of either "present" or "1". Target samples contain target bounding boxes. Empty target samples are image samples without an aircraft flight recorder target, and the target category label of empty target samples has a record attribute of either "absent" or "0". Empty target samples do not contain target bounding boxes.
[0012] Preferably, the aircraft flight recorder image samples in the aircraft flight recorder image sample dataset are obtained from images of aircraft wreckage collected by drones, handheld camera devices, vehicle-mounted camera devices, or rescue robots. Target samples include image samples of the complete target of the aircraft flight recorder, the partial target of the aircraft flight recorder, and / or closely related components of the aircraft flight recorder.
[0013] Preferably, the aircraft flight recorder image samples in the aircraft flight recorder image sample dataset undergo preprocessing including size normalization, image format conversion, and annotation format conversion. The enhancement processing of the aircraft flight recorder image samples in the aircraft flight recorder image sample dataset includes one or more of the following: random scaling, random cropping, brightness perturbation, contrast perturbation, rotation, translation, mosaic enhancement, noise perturbation, and occlusion simulation.
[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) The target detection and localization model of the present invention utilizes the SPD-Conv module to transfer spatial information to the channel dimension for downsampling through spatial-to-depth rearrangement and non-stepping convolution. This can fully preserve the early fine-grained features such as the edge, texture and local color of the flight recorder target, and avoid the loss of key details during downsampling, thereby improving the recognition effect of damaged, partially occluded or deformed flight recorder targets. At the same time, the LSKA attention mechanism is integrated into the SPPF module to obtain the SPPF-LSKA attention module. Based on the different receptive field context features obtained by multi-scale pooling, spatial and channel dual attention weights are generated through large kernel separable convolution to enhance the features of the key area where the flight recorder is located. This effectively improves the stability of the model in flight recorder target detection under complex background interference, local occlusion and target appearance damage conditions. A small target detection layer is also added, which can integrate the shallow details of the backbone network, enhance the perception ability of small-scale aircraft flight recorder targets in complex crash scene images, and effectively reduce the probability of missed detection.
[0015] (2) The target detection and localization model of this invention is based on the YOLOv8n network and improved by introducing a small target detection layer, SPD-Conv module and LSKA attention mechanism. The computational complexity remains low and the inference speed is still higher than 100 FPS, achieving a good balance between detection accuracy and inference efficiency. This invention can be deployed on rescue and search equipment such as inspection and rescue drones and rescue robots. It can quickly and accurately identify and locate flight data recorders, cockpit voice recorders and their visible components at aircraft accident sites, wreckage search sites or simulated crash scenarios, providing real-time auxiliary decision-making information for search and rescue personnel and greatly improving search efficiency. Attached Figure Description
[0016] Figure 1 This is a flowchart of the aircraft flight recorder detection and positioning method of the present invention; Figure 2 This is a structural principle diagram of the target detection and localization model in the embodiment; Figure 3 This is a schematic diagram of the SPD-Conv module in the embodiment; Figure 4 This is a schematic diagram of the SPPF-LSKA attention module in the embodiment; Figure 5 A schematic diagram illustrating the detection and positioning results of an aircraft flight recorder, as an example of an implementation. Figure 6 This is a performance comparison chart of the target detection and localization model of this invention and the existing YOLOv8n network when applied to target detection in an aircraft flight recorder. Detailed Implementation
[0017] The present invention will be further described in detail below with reference to embodiments: Example like Figures 1-4 As shown, a method for detecting and locating aircraft flight recorders based on image recognition deep learning is presented. The method includes: S1. Obtain the search image sequence dataset, which includes image data and the corresponding location information of the images. Construct a target detection and localization model based on an improved YOLOv8n network architecture. The target detection and localization model includes a backbone network, a neck feature fusion network, and a detection head. The target detection and localization model is trained using an aircraft flight recorder image sample dataset for target detection and recognition. Preferably, the aircraft flight recorder image sample dataset includes aircraft flight recorder image samples, target category labels, and target bounding boxes. The aircraft flight recorder image samples contain target samples and empty target samples. Target samples are image samples with an aircraft flight recorder target, and the target category label of a target sample is recorded as either "present" or "1". Target samples contain target bounding boxes. Empty target samples are image samples without an aircraft flight recorder target, and the target category label of an empty target sample is recorded as either "absent" or "0". Empty target samples do not contain target bounding boxes. The aircraft flight recorder image sample dataset consists of images of aircraft wreckage collected by drones, handheld cameras, vehicle-mounted cameras, or rescue robots. Targeted samples include images of the complete aircraft flight recorder target, images of parts of the aircraft flight recorder target, or / and closely related components of the aircraft flight recorder.
[0018] In some embodiments, the aircraft flight recorder image samples in the aircraft flight recorder image sample dataset undergo preprocessing including size normalization, image format conversion, and annotation format conversion. The enhancement processing of the aircraft flight recorder image samples in the aircraft flight recorder image sample dataset includes one or more of random scaling, random cropping, brightness perturbation, contrast perturbation, rotation, translation, mosaic enhancement, noise perturbation, and occlusion simulation, with the aim of enhancing the adaptability of the target detection and localization model to complex field environments.
[0019] In some embodiments, images of simulated aircraft accident scenes, wreckage search sites, or publicly available flight recorder-related images are acquired. These images may contain intact flight recorders, damaged flight recorders, partial shells, location beacon components, or other visible structures related to the flight recorder. For each image, a rectangular bounding box is used to label the target area of the flight recorder, and the category label is set to "flight_recorder". When it is necessary to distinguish the target state, multiple category labels such as intact flight recorder, damaged flight recorder, and flight recorder components can also be set. After labeling, the aircraft flight recorder image sample dataset is divided into training, validation, and test sets according to a preset ratio. For example, a 7:2:1 ratio can be used, or the ratio can be adjusted according to the number of samples and the application scenario. Before model training, the image size is normalized to 640×640, and random scaling, random rotation, horizontal flipping, brightness perturbation, contrast perturbation, mosaic enhancement, random occlusion, and noise perturbation are applied to simulate scale changes, occlusion, contamination, low light, and background interference in complex crash scenes.
[0020] S2. The image data from the search image sequence dataset are sequentially input into the target detection and localization model for recognition. The backbone network performs multi-scale feature extraction layer by layer on the image data, and then the SPPF-LSKA attention module performs attention weighting processing to output the weighted features. Figure 2 As shown, the backbone network of the preferred target detection and localization model of this invention includes a four-layer structure, and the upsampling branch and downsampling branch of the neck feature fusion network each include a three-layer structure. The detection head includes a small target detection head specifically adapted for target detection in aircraft flight recorders (i.e., Figure 2 (P2 small target detection layer).
[0021] like Figure 2As shown, the backbone network of the target detection and localization model has two SPD-Conv modules connected sequentially at the front end of its four-layer structure (the two SPD-Conv modules together constitute the improved convolutional module Conv). The SPD-Conv module includes a space-to-depth transformation layer and a non-staggered convolutional layer with a stride of 1. The space-to-depth transformation layer performs interval sampling and rearrangement of the input features (the input features of the first SPD-Conv module are the features extracted from the image data, and the input features of the second SPD-Conv module are the output features of the first SPD-Conv module) according to a preset scale factor s to obtain rearranged features. The non-staggered convolutional layer performs channel mapping on the rearranged features and outputs the features. For example... Figure 2 As shown, the features output by the second SPD-Conv module are input into the four-layer structure of the backbone network. The first three layers of the four-layer structure of the backbone network (such as...) Figure 2 As shown, each layer (including a basic convolutional module and a C2f residual module) sequentially outputs features A1 to A3 at three different scales (all output by the C2f residual module of each layer). The last layer of the four-layer backbone network contains an SPPF-LSKA attention module. The SPPF-LSKA attention module extracts global multi-scale contextual information, calculates spatial and channel dual attention weights, and performs attention weighting on the output features of the last layer of the four-layer backbone network to output the weighted feature A4. The four-layer backbone network sequentially outputs features at four scales from shallow to deep (i.e., features A1 to A4).
[0022] S3, the neck feature fusion network performs layer-by-layer feature fusion on the weighted feature A4 and the output features of the first three layers of the backbone network (i.e., features A3 to A1) through the upsampling branch, and then performs layer-by-layer feature fusion through the downsampling branch to obtain several levels of fused features. For example... Figure 2 As shown, the three-layer structure of the upsampling branch of the neck feature fusion network corresponds one-to-one with the first three layers of the four-layer structure of the backbone network. The three-layer structure of the upsampling branch sequentially upsamples and processes the data, then concatenates and fuses it with the corresponding layers of the backbone network to output three-layer scale features B2 to B4, as shown below. Figure 2 As shown, each layer of the upsampling branch includes an upsampling module Upsample, a stitching module Concat, and a residual module C2f. Each layer of the upsampling branch processes and outputs three-layer scale features B2, B3, and B4 respectively. The output feature B4 of the upsampling branch is input into the small target detection head of the detection head.
[0023] like Figure 2As shown, the three-layer structure of the downsampling branch corresponds one-to-one with the three-layer structure of the upsampling branch. Each layer of the downsampling branch includes a convolution module Conv, a concatenation module Concat, and a residual module C2f. The three-layer structure of the downsampling branch performs downsampling processing in sequence and is concatenated and fused with the corresponding layer of the upsampling branch to output features C2 to C4. The three-layer structure of the downsampling branch outputs features C2, C3, and C4 respectively through the residual module C2f.
[0024] In some embodiments, when the input image data size is 640×640, the feature map is further upsampled to 160×160 in the neck feature fusion network and concatenated with the shallow features of the backbone network. This concatenated feature map is then processed by the C2f module before being input into the new detection head. This detection head corresponds to a higher resolution feature map, which can retain more edge, texture, color, and local structural information of small targets in the flight recorder. The backbone network includes an SPD-Conv module. For the input feature map X, the SPD-Conv module first rearranges X into multiple sub-feature maps according to spatial location and concatenates them along the channel dimension, transferring spatial information to the channel dimension. Then, feature extraction and channel mapping are completed through convolution with a stride of 1. This structure achieves downsampling while avoiding the direct discarding of some pixel position information by ordinary strided convolution. The SPPF-LSKA attention module of this invention is obtained by integrating the LSKA attention mechanism into the SPPF module. Specifically, the SPPF module obtains contextual features under different receptive fields through multiple pooling, while the LSKA module generates attention responses through large kernel separable convolution, which enhances the flight recorder-related regions in the contextual features and suppresses complex background regions, thereby improving the model's ability to express the flight recorder target.
[0025] During model training, pre-trained weights from the object detection and localization model can be used as initial weights, or training can begin from scratch. The training process employs stochastic gradient descent or the Adam optimizer, setting the learning rate, batch size, and number of training epochs, and implementing an early stopping strategy based on the validation set detection accuracy. Training losses include class loss, bounding box regression loss, and distributed focus loss. After training, the model weights with the best performance on the validation set are saved as the object detection and localization model for the aircraft flight recorder.
[0026] S4. The detection head utilizes several levels of fused features to perform target detection and recognition of the aircraft flight recorder and outputs the target detection results. The target detection results include whether the aircraft flight recorder target exists and the target detection box. The detection head also includes three scale detection heads, which sequentially receive features C2 to C4 and perform aircraft flight recorder target detection and recognition processing adapted to scale changes. Preferably, the detection head adopts a decoupled detection structure and performs target category detection and target detection box regression processing. The target category detection is whether it is an aircraft flight recorder target. If the target category detection is an aircraft flight recorder target, the target detection box regression processing predicts the target detection box (see [link to relevant documentation]). Figure 5 After the target is detected and identified, a target detection box is displayed at the target location on the aircraft flight recorder.
[0027] To evaluate the trained object detection and localization model, this embodiment can introduce precision (P), recall (R), average precision (AP), mean average precision (mAP), number of model parameters (Params), model complexity (GFLOPs), and frames per second (FPS) as evaluation metrics, expressed as follows: , , , , TP represents the number of positive samples predicted as positive, FP represents the number of negative samples predicted as positive, and FN represents the number of positive samples predicted as negative. AP represents the area enclosed by the PR (precision-recall) curve, M represents the number of classes in the dataset, N(p) represents the total number of images processed, and T(p) represents the processing time for the images.
[0028] To verify the improvement effect of each improved module on the target detection and localization model of the aircraft flight recorder, four ablation experiments were designed. A small target detection layer (P2), SPD-Conv, and LSKA module were added to the target detection and localization model respectively. Each improved module was added to the target detection and localization model in turn, and the final experimental results are shown in Table 1.
[0029]
[0030] After adding a small target detection layer, the model can more fully integrate shallow detail information from the backbone network, improving the feature fusion capability of the neck network. Simultaneously, the added detection head enhances the ability to identify and locate small targets on the flight recorder. Compared to YOLOv8n, the model's precision, recall, and mAP@0.5 increased by 6.59, 10.64, and 10.23 percentage points, respectively, reaching 87.86%, 74.69%, and 83.30%. Building upon this, the introduction of SPD-Conv further reduced precision by 1.42 percentage points, but increased recall and mAP@0.5 by 6.79 and 4.56 percentage points, respectively, reaching 81.48% and 87.86%, while further reducing the model's parameter count to 2.92 × 10⁻⁶. 6 Finally, by incorporating the LSKA attention mechanism into SPPF, the precision and mAP@0.5 improved by 1.12 and 0.41 percentage points respectively compared to the previous step, while the recall decreased slightly by 1.22 percentage points. The final model achieved precision, recall, and mAP@0.5 of 87.56%, 80.26%, and 88.27%, respectively. Overall, compared to YOLOv8n, the improved target detection and localization model of this invention improved precision, recall, and mAP@0.5 by 6.29, 16.21, and 15.20 percentage points respectively, with only a 0.19 × 10⁻⁶ parameter increase. 6 Furthermore, the FPS of each model group was higher than 100, indicating that the improved method significantly improved the detection and localization performance of flight recorders while maintaining high inference speed and only slightly increasing model complexity. To visually demonstrate the performance difference between the improved target detection and localization model of this invention and the existing YOLOv8n network architecture, this embodiment compares four aspects: precision, recall, mAP@0.5, and mAP@0.5:0.95. Figure 6 As shown, the improved model has a faster convergence speed, and its overall performance curves are higher than those of YOLOv8n, indicating that the improved model has better comprehensive detection performance.
[0031] S5. Filter the image data in the image sequence dataset to find the target of the aircraft flight recorder. Store the image data and the corresponding target detection results. Extract the location information corresponding to the image data and parse the geographical location data of the aircraft flight recorder target.
[0032] In actual detection, a search image sequence dataset acquired via drones is input into the trained target detection and localization model. The input image (i.e., image data from the search image sequence dataset) is resized and then fed into the backbone network. First, the SPD-Conv module performs early downsampling and fine-grained feature preservation; subsequently, feature extraction modules such as C2f obtain feature maps at different levels. The neck feature fusion network upsamples, downsamples, and concatenates channels of the feature maps at different levels to form multi-scale fusion features. Specifically, this invention features a 160×160 small target branch that fuses shallow detail features for detecting small or distant flight recorder targets in the image; 80×80, 40×40, and 20×20 detection branches are used to handle medium-scale and large-scale targets, respectively. Deep features enter the SPPF-LSKA attention module, where SPPF first obtains multi-scale contextual representations, and then the LSKA attention mechanism enhances features related to the flight recorder's shape, color, and structure. The multi-scale detection head outputs candidate boxes, class probabilities, and confidence scores. Non-maximum suppression is applied to all candidate boxes, and detection results with a confidence level greater than a preset threshold are retained.
[0033] Finally, the detection system outputs the bounding box coordinates, category, and confidence score of the flight recorder in the image, and can overlay the detection results onto the original image for display (see [link]). Figure 5 After the target is detected by the aircraft flight recorder, a corresponding target detection box is displayed. Based on the pose or positioning information of the image acquisition device, the detection result can be mapped to the on-site search coordinates to assist search and rescue personnel in determining the geographical location information of the flight recorder.
[0034] 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 method for detecting and locating aircraft flight recorders based on image recognition deep learning, characterized in that: The methods include: S1. Obtain the search image sequence dataset, which includes image data and the location information corresponding to the images; A target detection and localization model based on an improved YOLOv8n network architecture is constructed. The target detection and localization model includes a backbone network, a neck feature fusion network, and a detection head. The target detection and localization model is trained using an aircraft flight recorder image sample dataset for target detection and recognition. S2. Input the image data of the search image sequence dataset into the target detection and localization model for recognition. The backbone network performs multi-scale feature extraction on the image data layer by layer, and then performs attention weighting processing through the SPPF-LSKA attention module to output the weighted features. S3. The neck feature fusion network performs layer-by-layer feature fusion on the weighted features and the backbone network output features through the upsampling branch, and then performs layer-by-layer feature fusion through the downsampling branch to obtain several levels of fused features. S4. The detection head uses several levels of fusion features to perform target detection and recognition of the aircraft flight recorder and outputs the target detection results.
2. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 1, characterized in that: The target detection result includes the presence or absence of an aircraft flight recorder target and the target detection bounding box, and also includes the following methods: S5. Filter the image data in the image sequence dataset to find the target of the aircraft flight recorder. Store the image data and the corresponding target detection results. Extract the location information corresponding to the image data and parse the geographical location data of the aircraft flight recorder target.
3. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 1, characterized in that: The backbone network of the target detection and localization model includes a four-layer structure, and the upsampling branch and downsampling branch of the neck feature fusion network each include a three-layer structure. The detection head includes a small target detection head specifically adapted for target detection in aircraft flight recorders.
4. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 3, characterized in that: The backbone network of the target detection and localization model has two SPD-Conv modules connected sequentially at the front end of its four-layer structure. The SPD-Conv module includes a spatial-to-depth transformation layer and a non-staggered convolutional layer with a stride of 1. The spatial-to-depth transformation layer performs interval sampling and rearrangement of the input features according to a preset scale factor s to obtain rearranged features. The non-staggered convolutional layer performs channel mapping on the rearranged features and outputs the features. In the four-layer structure of the backbone network, the first three layers sequentially output features A1 to A3 at three different scales. The last layer of the four-layer structure of the backbone network contains an SPPF-LSKA attention module. The SPPF-LSKA attention module extracts global multi-scale contextual information, calculates spatial and channel dual attention weights, and performs attention weight weighting on the output features of the last layer of the four-layer structure of the backbone network to output the weighted feature A4.
5. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 3, characterized in that: The three-layer structure of the upsampling branch of the neck feature fusion network corresponds one-to-one with the first three layers of the four-layer structure of the backbone network. The three-layer structure of the upsampling branch is upsampled sequentially and spliced and fused with the corresponding layer of the backbone network to output three-layer scale features B2 to B4. The output feature B4 of the upsampling branch is input into the small target detection head of the detection head.
6. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 5, characterized in that: The three-layer structure of the downsampling branch corresponds one-to-one with the three-layer structure of the upsampling branch. The three-layer structure of the downsampling branch performs downsampling processing in sequence and splices and fuses with the corresponding layer of the upsampling branch to output features C2 to C4. The detection head also includes three scale detection heads. The three scale detection heads receive features C2 to C4 in sequence and perform target detection and recognition processing of the aircraft flight recorder that adapts to scale changes.
7. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to any one of claims 1 to 3, characterized in that: The detection head adopts a decoupled detection structure and performs target category detection and target detection box regression processing. The target category detection is to determine whether it is an aircraft flight recorder target. If the target category is detected as an aircraft flight recorder target, the target detection box regression processing predicts the target detection box.
8. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 1, characterized in that: The aircraft flight recorder image sample dataset includes aircraft flight recorder image samples, target category labels, and target bounding boxes. The aircraft flight recorder image samples contain target samples and empty target samples. Target samples are image samples with an aircraft flight recorder target, and the target category label of target samples has the attribute of either "present" or "1". Target samples contain target bounding boxes. Empty target samples are image samples without an aircraft flight recorder target, and the target category label of empty target samples has the attribute of either "absent" or "0". Empty target samples do not contain target bounding boxes.
9. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 8, characterized in that: The aircraft flight recorder image sample dataset consists of images of aircraft wreckage collected by drones, handheld cameras, vehicle-mounted cameras, or rescue robots. Target samples include images of the complete target of the aircraft flight recorder, images of a portion of the target of the aircraft flight recorder, and / or images of closely related components of the aircraft flight recorder.
10. The method for detecting and locating aircraft flight recorders based on image recognition deep learning according to claim 8, characterized in that: The aircraft flight recorder image samples in the aircraft flight recorder image sample dataset undergo preprocessing including size normalization, image format conversion, and annotation format conversion. The enhancement processing of the aircraft flight recorder image samples in the aircraft flight recorder image sample dataset includes one or more of the following: random scaling, random cropping, brightness perturbation, contrast perturbation, rotation, translation, mosaic enhancement, noise perturbation, and occlusion simulation.