CN117315453B - Underwater small target detection method based on underwater sonar image - Google Patents
Underwater small target detection method based on underwater sonar image Download PDFInfo
- Publication number
- CN117315453B CN117315453B CN202311552299.7A CN202311552299A CN117315453B CN 117315453 B CN117315453 B CN 117315453B CN 202311552299 A CN202311552299 A CN 202311552299A CN 117315453 B CN117315453 B CN 117315453B
- Authority
- CN
- China
- Prior art keywords
- target detection
- module
- training
- visible light
- underwater
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 208
- 238000012549 training Methods 0.000 claims abstract description 90
- 238000013528 artificial neural network Methods 0.000 claims abstract description 70
- 238000012795 verification Methods 0.000 claims abstract description 50
- 238000001914 filtration Methods 0.000 claims abstract description 49
- 238000000034 method Methods 0.000 claims abstract description 13
- 230000008447 perception Effects 0.000 claims description 17
- 238000012545 processing Methods 0.000 claims description 17
- 230000004927 fusion Effects 0.000 claims description 10
- 238000013135 deep learning Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 3
- 230000008014 freezing Effects 0.000 claims description 3
- 238000007710 freezing Methods 0.000 claims description 3
- 238000004891 communication Methods 0.000 abstract description 2
- 230000005540 biological transmission Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 2
- 239000011229 interlayer Substances 0.000 description 2
- 239000010410 layer Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000007670 refining Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Landscapes
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
Abstract
The invention relates to the technical field of image communication, in particular to an underwater small target detection method based on an underwater sonar image, which comprises the following steps: building a target detection depth neural network, building a visible light target detection training set and a visible light target detection verification set, and performing a first-stage training; training a model; constructing a noisy visible light data set and an underwater sonar data set and performing a second stage of training; noise filtering training; and constructing an underwater sonar image target detection training set and a verification set, and performing training in a third stage to obtain a final target detection depth neural network. The method provided by the invention is suitable for accurately detecting the underwater small target, and the training detection speed is relatively high.
Description
Technical Field
The invention relates to the technical field of image communication, in particular to an underwater small target detection method based on an underwater sonar image.
Background
Sonar is currently the most commonly used and most critical device in underwater small target detection, because acoustic waves have smaller attenuation and dispersion in water and can propagate in a longer range than electromagnetic waves and light waves, so that the sonar has the capability of visualizing the underwater dynamic environment and is very suitable for detecting small objects protruding from the seabed. However, the resolution of the sonar image is lower than that of a common visible light image, the sonar image is more easily interfered by various sea clutter noise, the outline of the target in the image is blurred, and especially small targets such as mines can even change in shape greatly, so that the detection of the small targets based on the underwater sonar image is very challenging. In addition, in an underwater environment, the transmission rate of sonar data is severely limited, the sonar data is difficult to be quickly transmitted to a remote high-performance computing platform for real-time processing, and based on the problems, it is very necessary to design a real-time high-precision target detection depth neural network which can be deployed and applied to a small underwater unmanned aircraft to accurately detect underwater small targets.
Disclosure of Invention
The invention aims to solve the technical problem of providing an underwater small target detection method based on an underwater sonar image, which is used for training each module of a target detection depth neural network through three stages, so that parameters of each module are adjusted, and the obtained final target detection depth neural network can be suitable for accurate detection of underwater small targets.
The invention is realized by the following technical scheme:
an underwater small target detection method based on an underwater sonar image comprises the following steps:
s1: building a target detection depth neural network, building a visible light target detection training set and a visible light target detection verification set, and training in a first stage: the built target detection depth neural network comprises a main network module, an encoder module, a decoder module, an intersection ratio perception query selection module, a detection head network module, a noise filtering module and a feature map refinement module, wherein a training set in visible light target detection data sets is selected to train the target detection depth neural network in a first stage, and verification is carried out on a verification set in the visible light target detection data sets to obtain a visible light target detection training set and a visible light target detection verification set;
s2: model training: training all modules in the target detection depth neural network by using the visible light target detection training set and the visible light target detection verification set to obtain a trained target detection depth neural network;
s3: constructing a noisy visible light data set and an underwater sonar data set and performing a second stage of training: randomly selecting a plurality of pictures from the visible light target detection training set constructed in the step S1, adding random noise to form a plurality of visible light pictures with random noise, and then selecting a plurality of pictures from the established sonar data set and a plurality of visible light pictures with random noise to form a noise filtering training set together;
s4: noise filtering training: freezing parameters of other modules except the noise filtering module, and performing noise filtering training on the noise filtering module by using a noise filtering training set to obtain a trained noise filtering module;
s5: constructing an underwater sonar image target detection training set and a verification set and performing training in a third stage: dividing the underwater sonar image data set into an underwater sonar image training set and an underwater sonar image verification set, training the unfrozen modules except the main network module, the noise filtering module and the feature map refinement module on the underwater sonar image training set in a third stage, and fine-tuning parameters of the unfrozen modules to obtain a final target detection depth neural network.
The underwater small target detection method based on the underwater sonar image further comprises the following step S7: and (5) inputting the underwater sonar image verification set in the step (S5) into a final target detection depth neural network to obtain a detection result of the verification set, and evaluating the overall performance of the target detection depth neural network.
Preferably, the visible light target detection data set in step S1 is a COCO2017 target detection data set.
Preferably, in step S1, when the verification is performed on the verification set in the visible light target detection dataset, the evaluation criterion uses a standard cocoap metric.
Optimally, in the step S2, all modules in the target detection deep neural network are trained through a visible light target detection training set based on a hundred-degree PaddlePaddle deep learning framework.
Further, step S2, when training all modules in the target detection depth neural network by using the visible light target detection training set and the visible light target detection verification set, specifically includes the following steps:
s21: the feature images with the most semantic information in the feature images with different scales output by the main network module are sent to the editor module for feature refinement treatment, and then are respectively sent to the corresponding noise filtering module for denoising treatment as well as other feature images;
s22: sequentially carrying out pairwise fusion on the feature images subjected to denoising treatment by a feature image thinning module to obtain a plurality of feature images with different scales after treatment;
s23: sequentially expanding the plurality of processed feature graphs with different scales along the length-width dimension, and splicing along the channel dimension function to obtain a feature graph form which can be processed by the cross-ratio perception query selection module, and then sending the feature graph form into the cross-ratio perception query selection module;
s24: the cross-over ratio perception inquiry selection module screens out the data strip corresponding to the highest classification score, initializes the data strip to an inquiry object mode, and sends the data strip to the detection head network module after cross-attention processing is carried out on the data strip in the decoder module together with the feature map which can be processed by the cross-over ratio perception inquiry selection module and is obtained in the step S23;
s25: the detection head network module obtains a detection result about an interested target in the image through detection, then solves loss between the detection result and a real detection frame by using a bipartite graph matching method, and carries out parameter callback on the target detection depth neural network through the solved loss;
s26: and detecting the target detection depth neural network by using the visible light target detection verification set to obtain the trained target detection depth neural network.
Further, in step S25, the loss between the detection result and the real detection frame is solved by:;
wherein:representing predicted loss->Representing the target real result->Representing the predicted detection result,/->Representing the true category of the target->Representing a real detection frame->Representing predicted target category,/->The prediction detection box is represented by a frame,the sum of the cross-ratio loss representing the bounding box predictor and the binary cross entropy loss of the class predictor,representing trans-scale feature fusion and cross-ratio, +.>And represents the cross ratio loss of the prediction result of the boundary box.
The invention has the beneficial effects that:
the underwater small target detection method based on the underwater sonar image provided by the invention has the following advantages:
1. the method comprises the steps that a visible light target detection data set is used for training the whole target detection depth neural network in a first stage, a noisy visible light data set and an underwater sonar data set are assembled for training and noise filtering in a second stage, finally an underwater sonar image target detection training set and a verification set are assembled for training in a third stage, parameters are adjusted, and a final target detection depth neural network is obtained, so that the problems that the resolution of an underwater sonar image is lower than that of a common visible light image, the problem that the outline of the target in the image is fuzzy due to the fact that the interference of various sea clutter noise is easier to occur, the transmission rate of sonar data in an underwater environment is severely limited, and the problem that the sonar data is difficult to be quickly transmitted to a remote high-performance computing platform for real-time processing can be solved, and the accurate detection of the underwater small target can be realized;
2. the noise filtering module and the feature map refinement module are provided, the feature map which is output by the main network module and contains the most semantic information in a plurality of feature maps with different scales is firstly sent to the editor module for feature refinement treatment, then is respectively sent to the corresponding noise filtering module for denoising treatment as well as other feature maps, and then the feature maps after denoising treatment are sequentially subjected to pairwise fusion by the feature map refinement module to obtain a plurality of feature maps with different scales after treatment, so that semantic information can be more effectively applied, the calculated amount required by a target detection depth neural network is greatly reduced, and the noise filtering module is convenient for separating and treating the noise features and the image features of the feature map, so that the processing speed can be increased, and the noise filtering precision can be improved; the method is convenient for sequentially expanding a plurality of processed feature graphs with different dimensions along the length-width dimension, and is convenient for processing the feature graphs which can be processed by the cross-ratio perception query selection module after being spliced along the channel dimension function, so that the refinement degree difference between the feature graphs with different dimensions to be fused can be reduced, the influence of the inter-layer refinement degree difference of image noise and feature layers on the accuracy of the target detection depth neural network is reduced, the detection performance of the target detection depth neural network is further improved, and the accurate detection of underwater small targets is convenient to realize.
Drawings
FIG. 1 is a schematic flow chart of the method of the invention.
Fig. 2 is a schematic flow chart of training all modules in the target detection depth neural network by using the visible light target detection training set and the visible light target detection verification set.
Detailed Description
The invention discloses an underwater small target detection method based on an underwater sonar image, which is shown in a specific flow chart in fig. 1 and comprises the following steps:
s1: building a target detection depth neural network, building a visible light target detection training set and a visible light target detection verification set, and training in a first stage: the built target detection depth neural network comprises a main network module, an encoder module, a decoder module, an intersection ratio perception query selection module, a detection head network module, a noise filtering module and a feature map refinement module, wherein a training set in visible light target detection data sets is selected to train the target detection depth neural network in a first stage, and verification is carried out on a verification set in the visible light target detection data sets to obtain a visible light target detection training set and a visible light target detection verification set;
the visible light target detection dataset is preferably a COCO2017 target detection dataset, the size of an image of the COCO2017 dataset is 3×640×640, the size of an underwater sonar dataset used in the later stage is smaller and is 3×480×480, so that the image is required to be randomly cut into 3×480×480 when the COCO2017 dataset is used for training a target detection depth neural network in the first stage, a training set in the COCO2017 dataset comprises 118287 images, a verification set comprises 5000 images, various high-quality labels are provided, the targets of interest for detection in the image comprise 80 subclasses in total, the class id number is discontinuous and is 90 at most, and a backbone network module can select HGNet as a backbone network;
s2: model training: training all modules in the target detection depth neural network by using the visible light target detection training set and the visible light target detection verification set to obtain a trained target detection depth neural network; the first stage utilizes the visible light target data set to fully train the whole target detection depth neural network, so that the main network module and the feature map refinement module can obtain strong feature extraction, noise elimination capability and feature refinement fusion capability;
s3: constructing a noisy visible light data set and an underwater sonar data set and performing a second stage of training: randomly selecting a plurality of pictures from the visible light target detection training set constructed in the step S1, adding random noise to form a plurality of visible light pictures with random noise, and then selecting a plurality of pictures from the established sonar data set and a plurality of visible light pictures with random noise to form a noise filtering training set together;
s4: noise filtering training: freezing parameters of other modules except the noise filtering module, and performing noise filtering training on the noise filtering module by using a noise filtering training set to obtain a trained noise filtering module;
the second stage of training is to freeze parameters of other modules except the noise filtering module, and build images of visible light data sets with noise and underwater sonar images for training, so that the noise filtering module can obtain strong noise removing capability, and the noise filtering capability of the noise filtering module on different data sets is further improved;
s5: constructing an underwater sonar image target detection training set and a verification set and performing training in a third stage: dividing the underwater sonar image data set into an underwater sonar image training set and an underwater sonar image verification set, training the unfrozen modules except the main network module, the noise filtering module and the feature map refinement module on the underwater sonar image training set in a third stage, and fine-tuning parameters of the unfrozen modules to obtain a final target detection depth neural network.
The parameters of the trunk network, the noise filtering module and the feature map refining module are frozen in training in the third stage, the feature extraction capacity, the noise filtering capacity and the feature refining fusion capacity of the target detection depth neural network are reserved, the network parameters of the unfrozen module are adjusted by utilizing the complete underwater sonar image data set, the target detection network applicable to the underwater sonar image can be finally obtained, namely the final target detection depth neural network, the underwater sonar image is input into the final target detection depth neural network, and the detection result can be automatically output, so that the underwater small target can be accurately detected.
Therefore, the underwater small target detection method based on the underwater sonar image provided by the invention is characterized in that the visible light target detection data set is used for carrying out first-stage training on the whole target detection depth neural network, then the noisy visible light data set and the underwater sonar data set are assembled for carrying out second-stage training and noise filtering, finally the underwater sonar image target detection training set and the verification set are assembled, parameters of the trunk network module, the noise filtering module and the feature map refinement module are frozen, the underwater sonar image target detection training set is used for training the target detection depth neural network, the final target detection depth neural network obtained after the parameters of the unfreezing module are adjusted is solved, the problems that the resolution of the underwater sonar image is lower than that of a general visible light image, the target contour in the underwater sonar image is fuzzy due to various sea clutter noise interferences, the transmission rate of the underwater sonar data in the underwater environment is severely limited, the underwater sonar image is difficult to be quickly transmitted to a remote high-performance computing platform for real-time processing are solved, the final target detection depth neural network can be suitable for underwater images, the underwater sonar image is automatically input into the final target detection depth neural network, and the underwater small target detection result can be accurately achieved.
The underwater small target detection method based on the underwater sonar image further comprises the following step S7: and (5) inputting the underwater sonar image verification set in the step (S5) into a final target detection depth neural network to obtain a detection result of the verification set, and evaluating the overall performance of the target detection depth neural network. And during verification, inputting the image of the verification set into a final target detection depth neural network, comparing the output detection result with the label of the verification set, if the detection result is matched with the label of the verification set, indicating that the final target detection depth neural network meets the requirements, and if the detection result is not matched with the label of the verification set, indicating that the final target detection depth neural network does not meet the requirements.
Preferably, in step S1, when the verification is performed on the verification set in the visible light target detection dataset, the evaluation criterion uses a standard cocoap metric.
Optimally, in the step S2, all modules in the target detection deep neural network are trained through a visible light target detection training set based on a hundred-degree PaddlePaddle deep learning framework. The hundred-degree PaddlePaddle deep learning framework is a deep learning platform with preferential performance, flexibility and easiness in use, and covers the services and technologies in multiple fields such as searching, image recognition, speech semantic recognition understanding, machine translation, user portrayal and the like in the aspect of the deep learning framework. Because there are few underwater sonar image samples available for training the target detection model, the invention adopts a migration learning strategy to train the model, so that the model can obtain better results on an underwater sonar image data set.
Further, step S2 is shown in fig. 2, and specifically includes the following steps when training all modules in the target detection depth neural network by using the visible light target detection training set and the visible light target detection verification set:
s21: the feature images with the most semantic information in the feature images with different scales output by the main network module are sent to the editor module for feature refinement treatment, and then are respectively sent to the corresponding noise filtering module for denoising treatment as well as other feature images; when training all modules in the target detection depth neural network by utilizing the visible light target detection training set and the visible light target detection verification set, the main network module outputs a plurality of feature images with different scales, the feature images with different scales contain different information, the rest of the feature images with different scales contain more semantic information, the feature images with the largest semantic information in the plurality of feature images with different scales output by the main network module are sent to the editor module for feature refinement processing, and then are respectively sent to the corresponding noise filtering module for denoising processing as the other feature images, so that the semantic information can be more effectively applied, the calculation amount required by the target detection depth neural network is greatly reduced, the noise filtering module can conveniently separate the noise features and the image features of the feature images, the processing speed can be increased, and the noise filtering precision can be improved;
s22: sequentially carrying out pairwise fusion on the feature images subjected to denoising treatment by a feature image thinning module to obtain a plurality of feature images with different scales after treatment;
s23: sequentially expanding the plurality of processed feature graphs with different scales along the length-width dimension, and splicing along the channel dimension function to obtain a feature graph form which can be processed by the cross-ratio perception query selection module, and then sending the feature graph form into the cross-ratio perception query selection module;
s24: the cross-over ratio perception inquiry selection module screens out the data strip corresponding to the highest classification score, initializes the data strip to an inquiry object mode, and sends the data strip to the detection head network module after cross-attention processing is carried out on the data strip in the decoder module together with the feature map which can be processed by the cross-over ratio perception inquiry selection module and is obtained in the step S23;
s25: the detection head network module obtains a detection result about an interested target in the image through detection, then solves loss between the detection result and a real detection frame by using a bipartite graph matching method, and carries out parameter callback on the target detection depth neural network through the solved loss;
s26: and detecting the target detection depth neural network by using the visible light target detection verification set to obtain the trained target detection depth neural network.
According to the underwater small target detection method based on the underwater sonar image, the noise filtering module and the feature map refinement module are added in the target detection depth neural network, the feature map which is output by the main network module and contains the most semantic information in the feature maps with different dimensions is sent to the editor module for feature refinement processing, then the feature map is sent to the corresponding noise filtering module for denoising processing as well as other feature maps, the feature map after denoising processing is sequentially subjected to pairwise fusion through the feature map refinement module to obtain the feature maps with different dimensions after processing, the feature maps with different dimensions after processing are sequentially unfolded along the length-width dimension, and are sent to the cross-domain than the perception query selection module for processing after being processed along the channel dimension function, so that the difference of the degree of refinement among the feature maps with different dimensions to be fused can be reduced, the influence of the image noise and the difference of the degree of the inter-layer degree of the feature layer on the accuracy of the target detection depth neural network is reduced, the detection depth neural network is further refined, the underwater small target detection accuracy is conveniently realized, the accuracy of the underwater small target detection can be effectively detected, the required depth of the target detection is greatly is improved, the neural network is calculated, and the required detection speed is greatly is reduced.
In step S25, the loss between the detection result and the real detection frame is solved by the following equation:;
wherein:representing predicted loss->Representing the target real result->Representing the predicted detection result,/->Representing the true category of the target->Representing a real detection frame->Representing predicted target category,/->The prediction detection box is represented by a frame,the sum of the cross-ratio loss representing the bounding box predictor and the binary cross entropy loss of the class predictor,representing trans-scale feature fusion and cross-ratio, +.>The cross ratio loss of the prediction result of the boundary box is represented, so that the aim of accelerating model convergence more easily can be achieved.
In summary, according to the underwater small target detection method based on the underwater sonar image, each module of the target detection depth neural network is trained through three stages, parameters of each module are adjusted, and the obtained final target detection depth neural network can be suitable for accurate detection of underwater small targets and is high in training detection speed.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (6)
1. An underwater small target detection method based on an underwater sonar image is characterized by comprising the following steps of: the method comprises the following steps:
s1: building a target detection depth neural network, building a visible light target detection training set and a verification set, and training in the first stage: the built target detection depth neural network comprises a main network module, an encoder module, a decoder module, an intersection ratio perception query selection module, a detection head network module, a noise filtering module and a feature map refinement module, wherein a training set in visible light target detection data sets is selected to train the target detection depth neural network in a first stage, and verification is carried out on a verification set in the visible light target detection data sets to obtain a visible light target detection training set and a visible light target detection verification set;
s2: model training: training all modules in the target detection depth neural network by using a visible light target detection training set and a visible light target detection verification set to obtain a trained target detection depth neural network, and specifically comprises the following steps:
s21: the feature images with the most semantic information in a plurality of feature images with different scales output by a main network module are sent to an encoder module for feature refinement treatment, and then are respectively sent to a corresponding noise filtering module for denoising treatment as well as other feature images;
s22: sequentially carrying out pairwise fusion on the feature images subjected to denoising treatment by a feature image thinning module to obtain a plurality of feature images with different scales after treatment;
s23: sequentially expanding the plurality of processed feature graphs with different scales along the length-width dimension, and splicing along the channel dimension function to obtain a feature graph form which can be processed by the cross-ratio perception query selection module, and then sending the feature graph form into the cross-ratio perception query selection module;
s24: the cross-over ratio perception inquiry selection module screens out the data strip corresponding to the highest classification score, initializes the data strip to an inquiry object mode, and sends the data strip to the detection head network module after cross-attention processing is carried out on the data strip in the decoder module together with the feature map which can be processed by the cross-over ratio perception inquiry selection module and is obtained in the step S23;
s25: the detection head network module obtains a detection result about an interested target in the image through detection, then solves loss between the detection result and a real detection frame by using a bipartite graph matching method, and carries out parameter callback on the target detection depth neural network through the solved loss;
s26: detecting the target detection depth neural network by using a visible light target detection verification set to obtain a trained target detection depth neural network;
s3: constructing a noisy visible light data set and an underwater sonar data set and performing a second stage of training: randomly selecting a plurality of pictures from the visible light target detection training set constructed in the step S1, adding random noise to form a plurality of visible light pictures with random noise, and then selecting a plurality of pictures from the established underwater sonar data set and a plurality of visible light pictures with random noise to form a noise filtering training set together;
s4: noise filtering training: freezing parameters of other modules except the noise filtering module, and performing noise filtering training on the noise filtering module by using a noise filtering training set to obtain a trained noise filtering module;
s5: constructing an underwater sonar image target detection training set and a verification set and performing training in a third stage: dividing the underwater sonar image data set into an underwater sonar image training set and an underwater sonar image verification set, training the unfrozen modules except the main network module, the noise filtering module and the feature map refinement module on the underwater sonar image training set in a third stage, and fine-tuning parameters of the unfrozen modules to obtain a final target detection depth neural network.
2. The underwater small target detection method based on the underwater sonar image according to claim 1, wherein the method comprises the following steps: further comprising step S7: and (5) inputting the underwater sonar image verification set in the step (S5) into a final target detection depth neural network to obtain a detection result of the verification set, and evaluating the overall performance of the target detection depth neural network.
3. The underwater small target detection method based on the underwater sonar image according to claim 1, wherein the method comprises the following steps: the visible light target detection data set in step S1 is a COCO2017 target detection data set.
4. The underwater small target detection method based on the underwater sonar image according to claim 1, wherein the method comprises the following steps: in step S1, when verification is performed on the verification set in the visible light target detection dataset, the evaluation criterion uses a standard cocoap metric.
5. The underwater small target detection method based on the underwater sonar image according to claim 1, wherein the method comprises the following steps: in the step S2, all modules in the target detection deep neural network are trained through a visible light target detection training set based on the hundred-degree PaddlePaddle deep learning framework.
6. The underwater small target detection method based on the underwater sonar image according to claim 1, wherein the method comprises the following steps: in step S25, the loss between the detection result and the real detection frame is solved by the following equation:
wherein:representing predicted loss->Representing the target real result->Representing the predicted detection result,/->Representing the true category of the target->Representing a real detection frame->Representing predicted target category,/->Representing a predictive detection box,/->Sum of cross-ratio loss representing boundary box predictors and binary cross entropy loss of class predictors, < >>Representing trans-scale feature fusion and cross-ratio, +.>And represents the cross ratio loss of the prediction result of the boundary box.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311552299.7A CN117315453B (en) | 2023-11-21 | 2023-11-21 | Underwater small target detection method based on underwater sonar image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311552299.7A CN117315453B (en) | 2023-11-21 | 2023-11-21 | Underwater small target detection method based on underwater sonar image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117315453A CN117315453A (en) | 2023-12-29 |
CN117315453B true CN117315453B (en) | 2024-02-20 |
Family
ID=89281296
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311552299.7A Active CN117315453B (en) | 2023-11-21 | 2023-11-21 | Underwater small target detection method based on underwater sonar image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117315453B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN118154993A (en) * | 2024-05-09 | 2024-06-07 | 南昌工程学院 | Bimodal underwater dam crack detection method based on acousto-optic image fusion |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110795991A (en) * | 2019-09-11 | 2020-02-14 | 西安科技大学 | Mining locomotive pedestrian detection method based on multi-information fusion |
CN111209952A (en) * | 2020-01-03 | 2020-05-29 | 西安工业大学 | Underwater target detection method based on improved SSD and transfer learning |
CN114463620A (en) * | 2022-01-26 | 2022-05-10 | 自然资源部第三海洋研究所 | Side-scan sonar marine substrate identification method and device based on YOLOv5 |
CN114758237A (en) * | 2022-04-19 | 2022-07-15 | 哈尔滨工程大学 | Construction method, detection method and construction device of automatic water delivery tunnel defect identification model, computer and storage medium |
CN114842208A (en) * | 2022-06-08 | 2022-08-02 | 南昌大学 | Power grid harmful bird species target detection method based on deep learning |
CN115880495A (en) * | 2022-12-22 | 2023-03-31 | 上海交通大学 | Ship image target detection method and system under complex environment |
CN116721112A (en) * | 2023-08-10 | 2023-09-08 | 南开大学 | Underwater camouflage object image segmentation method based on double-branch decoder network |
CN116912673A (en) * | 2023-07-19 | 2023-10-20 | 河南工业大学 | Target detection method based on underwater optical image |
-
2023
- 2023-11-21 CN CN202311552299.7A patent/CN117315453B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110795991A (en) * | 2019-09-11 | 2020-02-14 | 西安科技大学 | Mining locomotive pedestrian detection method based on multi-information fusion |
CN111209952A (en) * | 2020-01-03 | 2020-05-29 | 西安工业大学 | Underwater target detection method based on improved SSD and transfer learning |
CN114463620A (en) * | 2022-01-26 | 2022-05-10 | 自然资源部第三海洋研究所 | Side-scan sonar marine substrate identification method and device based on YOLOv5 |
CN114758237A (en) * | 2022-04-19 | 2022-07-15 | 哈尔滨工程大学 | Construction method, detection method and construction device of automatic water delivery tunnel defect identification model, computer and storage medium |
CN114842208A (en) * | 2022-06-08 | 2022-08-02 | 南昌大学 | Power grid harmful bird species target detection method based on deep learning |
CN115880495A (en) * | 2022-12-22 | 2023-03-31 | 上海交通大学 | Ship image target detection method and system under complex environment |
CN116912673A (en) * | 2023-07-19 | 2023-10-20 | 河南工业大学 | Target detection method based on underwater optical image |
CN116721112A (en) * | 2023-08-10 | 2023-09-08 | 南开大学 | Underwater camouflage object image segmentation method based on double-branch decoder network |
Non-Patent Citations (2)
Title |
---|
一种基于深度卷积神经网络的水下光电图像质量优化方法;张清博;张晓晖;韩宏伟;;光学学报;20180627(第11期);全文 * |
基于侧扫声纳的人工鱼礁自动识别方法研究;沈蔚;马建国;张进;帅晨甫;管明雷;;海洋测绘;20191125(第06期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN117315453A (en) | 2023-12-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10885381B2 (en) | Ship detection method and system based on multidimensional scene features | |
CN108647585B (en) | Traffic identifier detection method based on multi-scale circulation attention network | |
CN111123257B (en) | Radar moving target multi-frame joint detection method based on graph space-time network | |
CN110084234B (en) | Sonar image target identification method based on example segmentation | |
Isa et al. | Optimizing the hyperparameter tuning of YOLOv5 for underwater detection | |
CN112766087A (en) | Optical remote sensing image ship detection method based on knowledge distillation | |
Roch et al. | Automated extraction of odontocete whistle contours | |
CN117315453B (en) | Underwater small target detection method based on underwater sonar image | |
CN108444447B (en) | Real-time autonomous detection method for fishing net in underwater obstacle avoidance system | |
CN109271856A (en) | Remote sensing image object detection method based on expansion residual error convolution | |
CN113052834B (en) | Pipeline defect detection method based on convolution neural network multi-scale features | |
CN111860175B (en) | Unmanned aerial vehicle image vehicle detection method and device based on lightweight network | |
CN115546555B (en) | Lightweight SAR target detection method based on hybrid characterization learning enhancement | |
CN111126278A (en) | Target detection model optimization and acceleration method for few-category scene | |
CN115393734A (en) | SAR image ship contour extraction method based on fast R-CNN and CV model combined method | |
CN110751157B (en) | Image significance segmentation and image significance model training method and device | |
CN110570361B (en) | Sonar image structured noise suppression method, system, device and storage medium | |
CN116580393A (en) | Method and device for identifying nonmetallic inclusion in steel based on improved YOLOv5 | |
CN109508639B (en) | Road scene semantic segmentation method based on multi-scale porous convolutional neural network | |
Zhou et al. | Underwater occlusion object recognition with fusion of significant environmental features | |
CN112556682B (en) | Automatic detection algorithm for underwater composite sensor target | |
Khan et al. | Underwater target detection using deep learning: methodologies, challenges, applications and future evolution | |
CN117745944A (en) | Pre-training model determining method, device, equipment and storage medium | |
CN112800932A (en) | Method for detecting obvious ship target in marine background and electronic equipment | |
CN113034398A (en) | Method and system for eliminating jelly effect in urban surveying and mapping based on artificial intelligence |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |