White dolphin recognition system based on YOLOv5 target detection model

Abstract

The invention relates to the technical field of computer vision and marine organism monitoring, and discloses a Chinese white dolphin recognition system based on a YOLOv5 target detection model, and the system comprises an image collection module; a preprocessing module; a YOLOv5 target detection model is established; a post-processing module; and an output module. According to the Chinese white dolphin recognition system based on the YOLOv5 target detection model, the structure of the YOLOv5 model is optimized for underwater complex environments (such as light attenuation, water turbidity and target shielding), the feature extraction capacity of the YOLOv5 model for the Chinese white dolphin is enhanced, false detection and missing detection are reduced, and the recognition precision of the Chinese white dolphin is improved; the advantages of a YOLOv5 lightweight network are utilized, low-delay and high-frame-rate real-time detection is realized on embedded equipment (such as an unmanned aerial vehicle and an underwater robot) or an edge computing terminal, the timeliness requirement of field ecological monitoring is met, and real-time monitoring and efficient identification are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and marine life monitoring, and in particular to a Chinese white dolphin recognition system based on the YOLOv5 target detection model. Background Art

[0002] The Chinese white dolphin, a member of the order Cetacea, is a flagship species found in coastal waters. A small-to-medium-sized cetacean, it holds significant ecological, scientific, and cultural value, earning it the nickname "the giant panda of the sea." Since 2017, it has been listed as a "vulnerable" marine mammal on the IUCN Red List of Threatened Species and a Class I protected species on China's National List of Key Protected Wildlife.

[0003] Traditional protection of Chinese white dolphins relies heavily on manual visual observation or acoustic sonar detection. With the development of deep learning technology, current target detection algorithms are roughly divided into two types: two-stage algorithms such as Faster-RCNN, and single-stage detection algorithms such as SSD and YOLO. Existing technologies have many problems, such as poor image quality due to underwater environmental interference, false detections, insufficient real-time performance, and limited sample size resulting in weak model generalization. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides a Chinese white dolphin recognition system based on the YOLOv5 target detection model, which solves the problems raised in the above background technology through deep adaptation of deep learning technology and ocean monitoring scenarios.

[0006] (2) Technical solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a Chinese white dolphin recognition system based on the YOLOv5 target detection model, comprising:

[0008] Image acquisition module;

[0009] Preprocessing module;

[0010] YOLOv5 object detection model;

[0011] Post-processing module;

[0012] Output module.

[0013] Preferably, the YOLOv5 target detection model includes:

[0014] Backbone network;

[0015] Feature Pyramid Network (FPN);

[0016] Prediction head (Head).

[0017] Preferably, the YOLOv5 target detection model further includes:

[0018] Attention mechanism module;

[0019] Adaptive anchor box generation module.

[0020] Preferably, the preprocessing module includes:

[0021] underwater image enhancement unit;

[0022] Data augmentation unit.

[0023] Preferably, the post-processing module includes:

[0024] Non-maximum suppression (NMS) unit;

[0025] Confidence threshold adjustment unit.

[0026] Preferably, the system is deployed on at least one of the following hardware platforms:

[0027] Embedded devices;

[0028] drones or underwater robots;

[0029] Cloud server.

[0030] A method for identifying Chinese white dolphins comprises the following steps:

[0031] Step 1: Acquire image data containing Chinese white dolphins through the image acquisition module;

[0032] Step 2: Preprocess the image data, including denoising, enhancement and standardization;

[0033] Step 3: Input the preprocessed image into the optimized YOLOv5 object detection model to identify Chinese white dolphins;

[0034] Step 4: Post-process the detection results and output the location and confidence information of the Chinese white dolphin;

[0035] Step 5: Visualize or transmit the recognition results to the monitoring platform.

[0036] Preferably, the training steps of the YOLOv5 target detection model include:

[0037] Step 1: Build a Chinese white dolphin image dataset and annotate its location and category;

[0038] Step 2: Use transfer learning technology to fine-tune the pre-trained YOLOv5 model;

[0039] Step 3: Improve model generalization ability through data enhancement and hyperparameter optimization.

[0040] Preferably, the storage medium stores a computer program, and when the computer program is executed by the processor, the method for identifying Chinese white dolphins according to claim 7 or 8 is implemented.

[0041] Preferably, it comprises a memory, a processor and a computer program stored in the memory, and when the processor executes the computer program, it implements the method for identifying Chinese white dolphins according to claim 7 or 8.

[0042] (3) Beneficial effects

[0043] Compared with the existing technology, the present invention provides a Chinese white dolphin recognition system based on the YOLOv5 target detection model, which has the following beneficial effects:

[0044] 1. This Chinese white dolphin recognition system, based on the YOLOv5 target detection model, optimizes the YOLOv5 model structure to address complex underwater environments (such as light attenuation, water turbidity, and target occlusion), enhancing its feature extraction capabilities for Chinese white dolphins, reducing false and missed detections, and improving Chinese white dolphin recognition accuracy.

[0045] Leveraging the advantages of YOLOv5's lightweight network, low-latency, high-frame-rate real-time detection is achieved on embedded devices (such as drones and underwater robots) or edge computing terminals, meeting the timeliness requirements of field ecological monitoring and enabling real-time monitoring and efficient identification.

[0046] Through data enhancement and transfer learning technology, the system can adapt to shooting conditions in different sea areas and seasons, improve its robustness in changing natural environments, and enhance the model's generalization ability;

[0047] Automated identification technology replaces traditional manual observation or passive acoustic monitoring, reducing manpower input and equipment maintenance costs. It is suitable for long-term, large-scale marine ecological surveys, reducing monitoring costs and labor dependence.

[0048] It provides accurate data support for scientific research work such as Chinese white dolphin population statistics, migration route analysis, and habitat protection. It can also be used for ship collision avoidance warning, reduce the interference of human activities on endangered species, and support ecological protection and scientific research applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic diagram of the overall technical architecture of the present invention;

[0050] Figure 2 Schematic diagram of the network structure of YOLOv5s of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] See also Figure 1-2 The present invention provides a technical solution: a Chinese white dolphin recognition system based on the YOLOv5 target detection model, comprising:

[0053] An image acquisition module, used to acquire underwater or surface image data containing Chinese white dolphins;

[0054] Preprocessing module, used to denoise, enhance and standardize image data;

[0055] The YOLOv5 object detection model is used to identify Chinese white dolphins from preprocessed images. The YOLOv5 object detection model is trained and optimized for the characteristics of Chinese white dolphins.

[0056] The post-processing module is used to filter and optimize the detection results and output the location and confidence information of the Chinese white dolphin;

[0057] The output module is used to visualize or transmit the recognition results to the monitoring platform.

[0058] In the present invention, the YOLOv5 target detection model includes:

[0059] The backbone network (Backbone) uses the CSPDarknet53 structure to extract multi-scale features of images;

[0060] Feature Pyramid Network (FPN), used to fuse features of different scales and enhance small object detection capabilities;

[0061] Prediction head (Head), used to output detection boxes and category probabilities.

[0062] In the present invention, the YOLOv5 target detection model further includes:

[0063] An attention mechanism module, embedded in the backbone network or feature pyramid network, is used to enhance the focus on key features of the Chinese white dolphin (such as dorsal fin and body color);

[0064] The adaptive anchor frame generation module is used to dynamically adjust the anchor frame size according to the morphological characteristics of the Chinese white dolphin.

[0065] In the present invention, the pre-processing module includes:

[0066] Underwater image enhancement unit for correcting color distortion and removing scattered noise;

[0067] The data augmentation unit is used to expand the training data by rotating, scaling, and adding noise.

[0068] In the present invention, the post-processing module includes:

[0069] Non-maximum suppression (NMS) unit, used to filter redundant detection boxes;

[0070] The confidence threshold adjustment unit is used to dynamically adjust the detection sensitivity according to environmental conditions.

[0071] In the present invention, the system is deployed on at least one of the following hardware platforms:

[0072] Embedded devices;

[0073] drones or underwater robots;

[0074] Cloud server.

[0075] A method for identifying Chinese white dolphins comprises the following steps:

[0076] Step 1: Acquire image data containing Chinese white dolphins through the image acquisition module;

[0077] Step 2: Preprocess the image data, including denoising, enhancement and standardization;

[0078] Step 3: Input the preprocessed image into the optimized YOLOv5 object detection model to identify Chinese white dolphins;

[0079] Step 4: Post-process the detection results and output the location and confidence information of the Chinese white dolphin;

[0080] Step 5: Visualize or transmit the recognition results to the monitoring platform.

[0081] In the present invention, the training steps of the YOLOv5 target detection model include:

[0082] Step 1: Build a Chinese white dolphin image dataset and annotate its location and category;

[0083] Step 2: Fine-tune the pre-trained YOLOv5 model using transfer learning technology. (Using transfer learning technology, fine-tune the YOLOv5 model pre-trained on a large-scale natural scene image dataset. This pre-trained model is suitable for Chinese white dolphin recognition scenarios because the marine environment frequented by Chinese white dolphins has certain similarities with natural scenes. The pre-trained model's learning results on features such as shape, texture, and color of objects in natural scenes help the model quickly focus on the characteristics of Chinese white dolphins, reducing the amount of data and time required for training, and improving model training efficiency and recognition accuracy.)

[0084] Step 3: Improve model generalization ability through data enhancement and hyperparameter optimization.

[0085] In the present invention, the storage medium stores a computer program, and when the computer program is executed by a processor, the method for identifying Chinese white dolphins according to claim 7 or 8 is implemented.

[0086] The present invention includes a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, a method for identifying Chinese white dolphins according to claim 7 or 8 is implemented.

[0087] 1. Analysis of technical issues

[0088] 1. Existing Chinese white dolphin monitoring technology has the following core pain points:

[0089] (1) Environmental interference: underwater light attenuation and obstruction by suspended objects lead to poor image quality;

[0090] (2) Missed detection of small targets: When photographed from a distance, white dolphins occupy only a very small area of the image (<5% pixels);

[0091] (3) Insufficient real-time performance: Traditional algorithms are difficult to implement real-time processing on embedded devices (<30FPS);

[0092] (4) Scarcity of samples: The scale of the labeled dataset is limited and the model generalization ability is weak.

[0093] 2. The overall technical framework of the present invention is as follows Figure 1 As shown in the figure, the system adopts the technical route of "end-to-end deep learning + lightweight deployment".

[0094] 2. Technical Effect

[0095] 1. In order to effectively evaluate the effectiveness of the target detection model, this paper selects precision P, recall R, mean average precision (mAP) and F1 score to evaluate the model performance.

[0096] Precision P refers to the percentage of the number of correctly predicted samples in the prediction data set to the number of positive samples predicted by the model.

[0097] Recall rate R (Recall) refers to the percentage of correctly predicted samples in the prediction data set to the actual number of positive samples.

[0098] The calculation of the above indicators is shown in the following formulas (1)-(4):

[0099]

[0100] Among them, TP is the number of correctly predicted positive samples, FP is the number of incorrectly predicted positive samples, FN is the number of actual positive samples predicted as negative samples, N is the total number of categories, and M is the total number of training BatchSizes.

[0101] 2. To evaluate the detection results of this invention, all models were trained on the same dataset using the same environment. The YOLOv5s model in this paper was compared with the mainstream one-stage network SSD and the two-stage network Faster-RCNN. The comparison results are shown in Table 1. YOLOv5s's F1 score is significantly higher than the other two models. Although its mAP is slightly lower than SSD and FastR-CNN, it has the highest accuracy and the smallest model size. Therefore, as a lightweight model, YOLOv5s has good model performance and is more suitable for deployment on other devices with limited resources.

[0102]

[0103]

[0104] Table 1 Comparative experiments of different target detection algorithms

[0105] 3. Brief Description of the Figures

[0106] 1. Chinese white dolphins primarily inhabit nearshore waters with depths of no more than 20 meters and within approximately 6 kilometers of shore. Photographing Chinese white dolphins in these waters is often affected by obstructions such as waves, and noise interference from factors such as weather and tides often results in suboptimal recognition. To accommodate images of varying sizes while ensuring both speed and accuracy, the research team selected YOLOv5s as the target detection model.

[0107] The network structure of YOLOv5s consists of input, backbone, neck and head (such as Figure 2YOLOv5s' input consists of three main components: Mosaic data augmentation, adaptive anchor box calculation, and adaptive image scaling. Mosaic data augmentation combines four images through random scaling, cropping, and arrangement to enrich the detection dataset. Adaptive anchor boxes calculate the optimal anchor box values for different datasets, reducing computational effort. Adaptive image scaling fixes images of varying sizes to 640 pixels by 640 pixels as input, improving object detection speed. The backbone network (Backbone) of YOLOv5s alternates between the CBS and C3 modules. The C3 module draws on the cross-stage partial network (CSPNet) structure to extract the features of Chinese white dolphins in images. The neck network combines the feature pyramid FPN structure and the path aggregation network PAN to fuse the feature information of Chinese white dolphins at different layers, enhance the positioning information of the underlying features, and pass it to the top layer. This retains multi-scale feature information, allowing the target detection algorithm to use targets of different sizes. At the output end, YOLOv5s uses downsampling to extract the feature layer from the backbone network, obtaining three feature layers to be predicted. The prediction results of different layers are then integrated, and the non-maximum suppression (NMS) algorithm is used to eliminate redundant boxes to obtain the final prediction result.

[0108] 4. System Hardware Composition

[0109] 1. The hardware implementation of this system mainly includes the following components:

[0110] (1) Image acquisition equipment: A DJI M300 RTK drone equipped with an H20T hybrid sensor camera, supporting 4K visible light shooting and 640×512 resolution thermal imaging;

[0111] (2) Edge computing device: NVIDIA Jetson AGX Xavier, equipped with 32GB memory and 512-core Volta GPU;

[0112] (3) Communication module: Huawei ME909s-8214G communication module;

[0113] (4) Power supply system: Adopts DJI TB60 intelligent battery pack, supporting dual battery redundant power supply.

[0114] 2. Software Architecture Implementation

[0115] The system software architecture is divided into three levels:

[0116] (1) Data collection layer:

[0117] Realize real-time video stream capture and build a video pipeline using the GStreamer framework

[0118] Support RTSP protocol transmission, delay controlled within 200ms

[0119] (2) Algorithm processing layer:

[0120] Implementing an improved YOLOv5 model based on the PyTorch 1.10 framework

[0121] (3) Application service layer:

[0122] Use Flask to build a RESTful API service.

[0123] 3. Improve the implementation of the YOLOv5 model

[0124] The model improvement includes the following key technical points:

[0125] (1) Attention mechanism module:

[0126] class CBAM(nn.Module):

[0127] def__init__(self,channels,reduction=16):

[0128] super(CBAM,self).__init__()

[0129] self.channel_attention=ChannelAttention(channels,reduction)

[0130] self.spatial_attention=SpatialAttention()

[0131] def forward(self,x):

[0132] x = self.channel_attention(x)

[0133] x = self.spatial_attention(x)

[0134] return x

[0135] (2) Adaptive anchor box calculation:

[0136] #Calculate dedicated anchor boxes based on the k-means++ algorithm

[0137] anchors = kmeans_anchors (dataset, n = 9, img_size = 640) # Get the optimized anchor box size

[0138] [(12,16),(30,36),(62,45),(85,68),(120,90)]

[0139] (3) Data enhancement strategy:

[0140] #Configure enhanced parameters in train.py

[0141] hyp={

[0142] 'lr0':0.01,

[0143] 'momentum':0.937,

[0144] 'weight_decay':0.0005,

[0145] 'fliplr':0.5,

[0146] 'mixup':0.1,

[0147] 'hsv_h':0.015,

[0148] 'hsv_s':0.7,

[0149] 'hsv_v':0.4,

[0150] 'degrees':15,

[0151] 'translate':0.1,

[0152] 'scale':0.5,

[0153] 'shear':0.0

[0154] 4. System workflow

[0155] (1) Initialization phase:

[0156] Load the pre-trained model weights, initialize the video capture interface, and establish a data connection with the cloud platform.

[0157] (2) Real-time detection stage:

[0158] while True:

[0159] frame = camera.read()#Get video frame

[0160] preprocessed = preprocess(frame) # Image preprocessing

[0161] detections = model (preprocessed) # target detection

[0162] results = postprocess(detections)#post-processing

[0163] send_to_cloud(results)#Result upload

[0164] (3)Exception handling mechanism:

[0165] Realize the function of automatically caching data when the network is disconnected.

[0166] 5. Performance optimization measures

[0167] (1) Model quantification:

[0168] python export.py--weights yolov5s.pt--include engine--half

[0169] (2) Multithreaded processing:

[0170] # Make Python threading module implement class VideoThread(threading.Thread):

[0171] def run(self):

[0172] while not self.stopped:

[0173] #Video capture and processing

[0174] (3) Memory optimization:

[0175] A circular buffer is used to manage video frames.

[0176] Implement zero-copy data transfer mechanism.

[0177] 6. Deployment and implementation plan

[0178] (1) UAV deployment:

[0179] Use Docker containers to package applications.

[0180] Configure the service to start automatically.

[0181] (2) Ground station configuration:

[0182] Build a receiving server cluster based on Kubernetes.

[0183] 7. System calibration and testing

[0184] (1) On-site calibration process:

[0185] 1. Collect test data under different lighting conditions

[0186] 2. Adjust model parameter thresholds

[0187] 3. Verify detection accuracy

[0188] (2) Performance test indicators:

[0189] Average processing latency on Jetson AGX Xavier: 22ms

[0190] Model memory usage: 1.2GB

[0191] Continuous working stability: >72 hours without failure

[0192] This specific implementation can be adjusted according to the needs of actual application scenarios, and all equivalent variations or replacements based on this technical solution should be included in the scope of protection of the present invention.

[0193] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A Chinese white dolphin recognition system based on the YOLOv5 target detection model, characterized in that: include: Image acquisition module; Preprocessing module; YOLOv5 object detection model; Post-processing module; The output module is used to visualize or transmit the recognition results to the monitoring platform.

2. The Chinese white dolphin recognition system based on the YOLOv5 target detection model according to claim 1, characterized in that: The YOLOv5 target detection model includes: Backbone network; Feature Pyramid Network (FPN); Prediction head (Head).

3. The Chinese white dolphin recognition system based on the YOLOv5 target detection model according to claim 1, characterized in that: The YOLOv5 target detection model further includes: Attention mechanism module; Adaptive anchor box generation module.

4. The Chinese white dolphin recognition system based on the YOLOv5 target detection model according to claim 1, characterized in that: The pre-processing module comprises: underwater image enhancement unit; Data augmentation unit.

5. The Chinese white dolphin recognition system based on the YOLOv5 target detection model according to claim 1, characterized in that: The post-processing module includes: Non-maximum suppression (NMS) unit; Confidence threshold adjustment unit.

6. The Chinese white dolphin recognition system based on the YOLOv5 target detection model according to claim 1, characterized in that: The system is deployed on at least one of the following hardware platforms: Embedded devices; drones or underwater robots; Cloud server.

7. A method for identifying Chinese white dolphins, characterized in that: The following steps are involved: Step 1: Acquire image data containing Chinese white dolphins through the image acquisition module; Step 2: Preprocess the image data, including denoising, enhancement and standardization; Step 3: Input the preprocessed image into the optimized YOLOv5 object detection model to identify Chinese white dolphins; Step 4: Post-process the detection results and output the location and confidence information of the Chinese white dolphin; Step 5: Visualize or transmit the recognition results to the monitoring platform.

8. The method for identifying Chinese white dolphins according to claim 7, characterized in that: The training steps of the YOLOv5 target detection model include: Step 1: Build a Chinese white dolphin image dataset and annotate its location and category; Step 2: Use transfer learning technology to fine-tune the pre-trained YOLOv5 model; Step 3: Improve model generalization ability through data enhancement and hyperparameter optimization.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which, when executed by a processor, implements the method for identifying Chinese white dolphins according to claim 7 or 8.

10. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory, wherein when the processor executes the computer program, the method for identifying Chinese white dolphins according to claim 7 or 8 is implemented.

Images