CN117830813B - Small celestial body surface rock detection method - Google Patents

Small celestial body surface rock detection method Download PDF

Info

Publication number
CN117830813B
CN117830813B CN202410085265.XA CN202410085265A CN117830813B CN 117830813 B CN117830813 B CN 117830813B CN 202410085265 A CN202410085265 A CN 202410085265A CN 117830813 B CN117830813 B CN 117830813B
Authority
CN
China
Prior art keywords
celestial body
detection
rock
body surface
small
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
Application number
CN202410085265.XA
Other languages
Chinese (zh)
Other versions
CN117830813A (en
Inventor
王彬
冯哲
黄鹏程
米布武
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Kunming University of Science and Technology
Original Assignee
Kunming University of Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Kunming University of Science and Technology filed Critical Kunming University of Science and Technology
Priority to CN202410085265.XA priority Critical patent/CN117830813B/en
Publication of CN117830813A publication Critical patent/CN117830813A/en
Application granted granted Critical
Publication of CN117830813B publication Critical patent/CN117830813B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/54—Extraction of image or video features relating to texture
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a method for detecting rocks on the surface of a celestial body, which comprises the following steps: constructing a small celestial body surface rock detection reference database; scaling pictures in the small celestial body surface rock detection database to a preset specification, sending the pictures into a backbone network to extract all levels of features, and finally enhancing rock features in the small celestial body feature map through a multi-head self-attention module MHSA in the backbone network, and then obtaining a small celestial body surface picture feature map through SPPF operation; inputting the small celestial body surface feature map to Neck networks of the networks, adding small target detection layers through Neck networks, fusing high-level semantic features in the low-level feature map, fusing low-level information and texture features in the high-level feature map, and performing multi-scale fusion on the obtained feature maps to obtain a multi-scale small celestial body surface feature map to be detected; and sending the multi-scale surface picture feature map of the small celestial body to be detected into a detection head detect, and predicting the position coordinates and the confidence of the rock in the map. Compared with the prior art, the invention has high identification accuracy and recall rate, thereby being capable of better completing the detection task of the celestial body rock.

Description

Small celestial body surface rock detection method
Technical Field
The invention relates to a method for detecting rocks on the surface of a celestial body, and belongs to the field of deep space exploration.
Background
The detection of the celestial bodies plays an indispensable role in revealing the puzzles of the birth and evolution of the solar system, and is an important technical approach for researching the formation process of the universe origin and the earth for human beings. The small celestial body is small in size, weak in attraction and more complex in surface topography and topography, and meteorites, rocks and steep slopes are distributed more, so that safe and accurate landing of the deep space detector faces great challenges, and meanwhile, higher requirements are provided for selection of landing areas. In order to improve the success rate of the detector for executing the safe landing or sampling return task, a safe area for the detector to land must be found in the process of the detector making a round-the-fly tour and descending on the small day, so that the landing risk is reduced. Rock is one of the most common topographical features of celestial surfaces, the primary object for landing site obstacle avoidance, and is also commonly used as a navigation "landmark" in optical navigation during probe landing. Therefore, it is necessary to develop research on surface rock target detection problems during the landing of the celestial body.
In recent years, many research students at home and abroad have studied rock detection on the surface of an extraterrestrial celestial body, a traditional rock detection algorithm generally adopts a method based on edge information and a method based on region information, two rock detection algorithms Rockster[1](Castano R,Estlin T,Gaines D,et al.Onboard autonomous rover science[C]//2007IEEEAerospace Conference.IEEE,2007:1-13.) and rockfinder[2](Castano A,Anderson R C,Castano R,et al.Intensity-based rock detection for acquiring onboard rover science[C]//Lunar and Planetary Science Conference.2004:2015.); based on extraction of edge information are proposed in the American JPL (Jet Propulsion Laboratory) laboratory, a new technology for detecting rock in a landing zone by only using a single image is provided in the prior art [3](Bajracharya M.Single image based hazard detection for a planetary lander[C]//Proceedings ofthe 5thbiannual WorldAutomation Congress.IEEE,2002,14:585-590.),Bajrachary, and a segmentation algorithm based on local intensity clustering is adopted to generate a rock size and abundance map according to shadow size and sun angle; prior art [4] (Zhang Zexu, cui Pingyuan. Method for detecting and evading rock for soft planet landing based on CCD landing camera [ J ]. Aviation journal, 2008, no.219 (06): 1510-1516.), zhang Zexu et al propose a method for detecting rock for soft planet landing based on CCD landing camera, which uses multi-threshold segmentation to perform rock detection in combination with C-means clustering method for rock target recognition; prior art [5] (Ding Meng, cao Yunfeng, wu Qingxian. Rock detection during passive image-based detector landing [ J ]. Photoengineering 2009,36 (01): 82-87.), ding Meng et al propose a rock area range detection technique based on shadows and contours. The algorithm belongs to a rock detection algorithm based on the traditional method, has a good detection effect, but has a low detection speed, has certain requirements on the shape and size of the rock, and cannot effectively detect small-size rocks with irregular shapes in a dark and weak environment. In addition, for the problems of sun illumination change, shielding caused by rugged ground surface shape and the like, false recognition and missing recognition are easy to generate, so that the recognition rate is low.
However, many rock features are not obvious due to the dark and weak environment on the celestial body; the small-size rocks on the celestial body are more and densely distributed, so that the difficulty of meteorite crater detection is further increased, the problem of missing detection and false detection of many rocks is caused, and a great challenge is brought to a rock detection task.
Disclosure of Invention
The invention provides a method for detecting rocks on the surface of a celestial body, which is used for solving the detection problem of a large number of densely distributed rocks, especially small-size rocks, aiming at the dark and weak environment on the celestial body.
The technical scheme of the invention is as follows:
a method for detecting rock on the surface of a celestial body, comprising:
S1, constructing a small celestial body surface rock detection reference database;
S2, zooming the pictures in the small celestial body surface rock detection database constructed in the S1 to a preset specification, then sending the pictures into a backbone network to extract all levels of features, and finally enhancing the rock features in the small celestial body feature map through a multi-head self-attention module MHSA in the backbone network, and then obtaining a small celestial body surface picture feature map through SPPF operation;
S3, inputting the small celestial body surface feature map output in the S2 to a Neck network of the network, fusing high-level semantic features in the low-level feature map, fusing low-level position information and texture features in the high-level feature map through a small target detection layer added to the Neck network, and performing multi-scale fusion on the feature maps of each layer obtained in the S2 to obtain a multi-scale small celestial body surface feature map to be detected;
and S4, sending the multi-scale surface picture feature map of the small celestial body to be detected obtained in the S3 into a detection head detect, and predicting the position coordinates and the confidence of the rock in the map.
The backbone network in S2 takes the backbone network of yolov as a framework, and MHSA operation is added before SPPF operation of the backbone network.
The Neck network in the S3 adopts a PANet structure, two characteristic transmission channels are respectively a deep-to-shallow transmission channel and a shallow-to-deep transmission channel, a small target detection layer is added on the PANet structure, the small target detection layer is used for upsampling the P3 level in the deep-to-shallow transmission channel to the P2 level, the characteristic map and a P2 level characteristic map concat output by a first C2f module in a backbone network are output by the first C2f module for detecting rocks of a first preset size, the first-scale small celestial body surface characteristic map to be detected is downsampled by the CBS module in the PANet transmission channel from the shallow-to-deep layer, and then is sent to the first C2f module in the shallow-to-deep transmission channel for detecting rocks of a second preset size; a second C2f module in the shallow-layer-to-deep-layer transmission channel outputs a third-scale celestial body surface characteristic map to be detected for detecting rocks of a third preset size; and a third C2f module in the shallow-layer-to-deep-layer transmission channel outputs a fourth-scale small celestial body surface characteristic map to be detected for detecting rocks of a fourth preset size.
In the step S4, the multi-scale surface feature map of the celestial body to be detected is sent to a head detection network detect for detection, the detection heads have four scales, each detection head has a decoupling structure, a classification task and a prediction frame position prediction task are respectively completed by different structures, regression parts of the four detection heads respectively predict the prediction frame positions, each prediction frame parameter comprises a center point x coordinate, a center point y coordinate, a width w and a height h, classification parts of the four detection heads predict n confidence degrees for each prediction frame, and n is the number of prediction categories; and finally mapping the prediction frames to the original image, framing out a target object on the original image, and finally obtaining a detection result.
The beneficial effects of the invention are as follows: aiming at the problems that the environment on the celestial body is dark and weak and a large amount of densely distributed rocks exist, particularly small-size rocks, the invention provides a celestial body surface rock detection method which has high identification accuracy and recall rate compared with the prior art, so that the celestial body rock detection task can be completed better. Specifically, compared with the traditional deep learning target detection algorithm, the multi-head self-attention module MHSA is added in the backbone network, so that the attention to rock characteristics is increased in the training process of the network, and the attention to flat ground characteristics is reduced, thereby improving the extraction capability of the network to the rock characteristics in a dark and weak environment, and solving the problem of poor detection effect caused by invalid and interference information in the surface image of the celestial body. The small target detection layer is added on the yolov basic model, a detection head capable of inputting a feature map with the size of 160 multiplied by 160 is added, small rocks with the height and width of 4 can be detected, and the detection capability of the small targets of the model is improved.
Drawings
FIG. 1 is a flow chart of the detection of rock on the surface of a celestial body according to an embodiment of the invention;
FIG. 2 is a schematic diagram illustrating the results of various data enhancement methods according to embodiments of the present invention;
FIG. 3 is a schematic diagram of a multi-headed self-attention module according to an embodiment of the present invention;
FIG. 4 is a diagram of a neck network architecture of a small object detection layer according to an embodiment of the present invention;
fig. 5 is a diagram of a complete network structure of the method according to the embodiment of the present invention.
Detailed Description
The invention will be further described with reference to the drawings and examples, but the invention is not limited to the scope.
Example 1: as shown in fig. 1-5, a method for detecting rocks on a celestial body surface includes:
S1, constructing a small celestial body surface rock detection reference database;
S2, enhancing the pictures in the small celestial body surface rock detection database constructed in the S1 by utilizing a plurality of online image enhancement modes, finally scaling to a preset specification, then sending the pictures into a backbone network (backbone network) to extract all levels of characteristics, enhancing the rock characteristics in the small celestial body characteristic map through a multi-head self-attention module MHSA at last in the backbone network, and obtaining a small celestial body surface picture characteristic map through SPPF operation; the on-line enhancement mode can improve the quality of the image and enhance the characteristics of the image; the goals of image enhancement may include improving contrast of the image, enhancing details, removing noise, adjusting brightness and color, etc.
S3, inputting the small celestial body surface feature map output in the S2 to a Neck network of the network, fusing high-level semantic features in the low-level feature map, fusing low-level position information and texture features in the high-level feature map through a small target detection layer added to the Neck network, and performing multi-scale fusion on the feature maps of each layer obtained in the S2 to obtain a multi-scale small celestial body surface feature map to be detected;
S4, sending the multi-scale surface picture feature mAP of the small celestial body to be detected obtained in the S3 into a detection head detect, predicting the position coordinates and the confidence coefficient of the rock in the mAP, calculating true labels to calculate evaluation indexes such as the accuracy P, the recall rate R, the average precision mAP and the like of detection results, and measuring the rock detection performance of the method.
Further, the backbone network in S2 takes the backbone network of yolov as a framework, and MHSA operations are added before SPPF operation of the backbone network.
Further, the Neck network in S3 adopts a PANet structure, two characteristic transmission channels are respectively a deep-to-shallow transmission channel and a shallow-to-deep transmission channel, a small target detection layer is added on the PANet structure, the small target detection layer is used for up-sampling the P3 level in the deep-to-shallow transmission channel to the P2 level, the characteristic map and a P2 level characteristic map concat output by a first C2f module in a backbone network are then processed by the first C2f module to obtain a first-scale small celestial body surface characteristic map to be detected for detecting rocks with a first preset size, the first-scale small celestial body surface characteristic map to be detected is down-sampled by the CBS module in the PANet transmission channel, and then is fed into the first C2f module in the deep-to-deep transmission channel to obtain a second-scale small celestial body surface characteristic map to be detected for detecting rocks with a second preset size; a second C2f module in the shallow-layer-to-deep-layer transmission channel obtains a third-scale small celestial body surface characteristic map to be detected and is used for detecting rocks with a third preset size; and a third C2f module in the shallow-layer-to-deep-layer transmission channel obtains a fourth-scale small celestial body surface characteristic map to be detected and is used for detecting rocks with a fourth preset size.
Further, in the step S4, the multi-scale surface feature map of the celestial body to be detected is sent to a head detection network detect for detection, the detection heads have four scales, each detection head has a decoupling structure, the classification task and the prediction frame position prediction task are respectively completed by different structures, the regression parts of the four detection heads respectively predict the prediction frame positions, each prediction frame parameter comprises a center point x coordinate, a center point y coordinate, a width w and a height h, the classification parts of the four detection heads predict n confidence degrees for each prediction frame, and n is the number of prediction categories; and finally mapping the prediction frames to the original image, framing out a target object on the original image, and finally obtaining a detection result.
Example 2 the following is an alternative embodiment of the invention, in conjunction with experimental data, as follows:
a method for detecting rock on the surface of a celestial body, comprising:
s1, constructing a small celestial body surface rock detection reference database;
S2, enhancing the pictures in the asteroid rock detection database constructed in the step S1 by utilizing a plurality of online image enhancement modes, finally scaling to 640 multiplied by 3, then sending the pictures into a backbone network to extract all levels of characteristics, enhancing the rock characteristics in the celestial body characteristic map through a multi-head self-attention module MHSA at last of the backbone network, and obtaining a celestial body surface picture characteristic map through SPPF operation;
As shown in fig. 3, the operations specifically performed to input the feature map into the multi-head self-attention module MHSA are as follows: the input feature map is segmented into a plurality of patches containing rock blocks at different positions, feature vectors of the patches of the different rock blocks are respectively extracted, each patch extracts three feature vectors of Q (query), K (key) and V (value), Q represents the attention degree of the topography (rock block or flat ground) of the current patch to the topography of other patches, K represents the importance degree of the topography of the other patches at the current position to the patch, Q and K are used for calculating the similarity of the rock blocks of the different patches, V is the feature expression after encoding the topography in each patch, and the topography feature vector output after the global topography feature is integrated in MHSA is related to the V of each patch. And a MHSA module obtains attention scores of the current rock patch on other patches by calculating the similarity between the feature vectors of the different rock patches. And it not only considers the similarity between the contents of different rock blocks, but also expresses the relative position relation of different rock blocks by giving relative position codes Rw and Rh to each block to form a relative position matrix R, and distinguishes the rock blocks by fusing the position information of each rock block in Attention score Attention. Q, K, V, R, attention is calculated as follows:
Q＝XWQ；
K＝XWK；
V＝XWV；
R＝Rw+Rh；
Step S3: neg neck network: the Neck network adopts PANet structure, and PANet has two characteristic transmission channels, namely deep to shallow transmission channel and shallow to deep transmission channel, the method adds small target detection layer in PANet structure, samples P3 level to P2 level in deep to shallow transmission channel, and adds the characteristic map to P2 level characteristic map concat (spliced in channel direction) which contains abundant positioning and texture information and is output by first C2f module in backbone network, then passes through a C2f module, the specification of the C2f output characteristic map is 160×160×256, these feature maps can detect small target rocks with height and width of 4 (4=640++160), the above 160×160×256 feature maps are downsampled to 80×80 by CBS module in the transmission channel from shallow to deep in PANet, and then sent to a C2f module after output concat with the second C2f in the transmission channel from deep to shallow to obtain feature maps with a specification of 80×80×256, which can detect rocks with height and width of 8, and the underlying rock positioning information and fine granularity features are propagated to the high layer (middle and large rock mass detection layer). PANet is that the transmission channel from the deep layer to the shallow layer transmits high-level semantic information of terrains such as rock, the transmission channel from the shallow layer to the deep layer transmits texture information and position information of terrains such as rock, the transmission channel from the shallow layer to the deep layer provides a channel for transmitting shallow texture and position information with a greatly reduced transmission distance, and a neck network structure diagram of the newly added small target detection layer is shown in fig. 4.
Step S4: and (3) sending the four-scale feature images obtained in the step (S3) into a head detection network detect for detection, wherein the four scales are shared by the detection heads, and the feature images of 160×160, 80×80, 40×40 and 20×20 are respectively input. Each detection head is of a decoupling structure, a classification task and a prediction frame position prediction task (regression task) are respectively completed by different structures, regression parts of the four detection heads respectively predict 160×160, 80×80, 40×40 and 20×20 prediction frames, each prediction frame parameter comprises four parts of a center point x coordinate, a center point y coordinate, a width w and a height h, the classification part predicts n confidence degrees for each prediction frame, and n is the number of prediction categories. And finally mapping the prediction frames to the original image, framing out a target object on the original image, and finally obtaining a detection result.
The complete network structure diagram of the method is shown in fig. 5, and the network is used as a small celestial body surface rock detection model; the small celestial body surface rock detection model takes yolov as a frame, MHSA operation is added before SPPF operation of a backbone network, and a small target detection layer is added in a neck network; the small target detection layer comprises Upsample, concat, C f and CBS which are sequentially connected, upsample of the small target detection layer is connected with the output of the last C2f in a shallow propagation path from the deep layer in the yolov neck network, concat of the small target detection layer is also used for being connected with the output of the first C2f in the yolov backbone network, the output of the C2f of the small target detection layer is also used as the input of the first detection head of the head network, and the output of the CBS of the small target detection layer is connected with the first Concat in the shallow to deep propagation path in the yolov neck network.
The loss function of the method of the invention is composed of two parts: the loss of the classification section and the loss of the bounding box regression.
The loss function of the classification part uses a two-class cross entropy loss function BCEloss, and the loss functions of the regression part are DFL loss (DistributionFocalLoss) and CIOUloss, which are formulated as follows:
BCE(pt,target）＝(target×log（pt）+(1-target)×log(1-pt))；
DFL(Si,Si+1＝-（yi+1-y）log Si+(y-yi)logSi+1)；
LossCIOU＝1-CIOU；
Losstotal＝LossCIOU+BCE+DFL；
Where pt is confidence of the prediction frame, target is a real label, y is one of four positioning parameters of the prediction frame, y i is a maximum integer smaller than y, y i+1 is a minimum integer larger than y, S i is probability of y i points in the distribution of prediction output, S i+1 is probability of y i+1 points in the predicted probability distribution, b is a prediction frame center point, b gt is a real frame center point, and ρ 2(b,bgt) is euclidean distance between the prediction frame center point and the real frame center point. c represents the diagonal distance of the minimum closure region that can contain both the predicted and real frames. w gt and h gt are the width and height of the real box and w and h are the width and height of the predicted box. IOU is the intersection ratio of the predicted and real frames.
Image detection accuracy (P), recall (R), F1 score, average accuracy (mAP), and number of frames per Second (FRAMES PER seconds, FPS) are introduced as evaluation indexes for the rock detection model, where: TP, FP, FS are the number of correct rock targets detected, the number of non-rock targets detected (the number of other targets that are false detected as rock targets), the number of rock targets not detected, respectively.
The specific application can be as follows:
In step S1, selecting a second detector tenon (Hayabusa 2) emitted in 2014 by japan JAXA (Japan Aerospace ExplorationAgency), taking a photo by an Optical Navigation Camera (ONC) when visiting the asteroid palace with the number 1999JU3, selecting 280 pictures in total from a Planetary Data System (PDS) of the national aviation and aerospace agency (NASA), expanding the pictures by 10 times by an offline data enhancement method of rotation (adopting 6 rotation operations with different angles in the invention), translation, high-S noise, random clipping and random brightness change, constructing a small celestial body surface rock detection reference database by 3080 pictures in total, marking rock targets on the pictures by using open source target detection marking software based on expert experience, saving the marking files in json format, then converting the marking files in json format into txt format used by yolo algorithm by using the python script, and finally, according to 8:1:1 into training, validation and test sets.
Inputting the pictures in the training set into the small celestial body surface rock detection model, scaling the pictures in the training set to 640 multiplied by 3 by an online image enhancement method to serve as the input of a backbone network of the small celestial body surface rock detection model, outputting a prediction frame and a category confidence coefficient after calculation of the small celestial body surface rock detection model, calculating loss with labels in the training set and back-propagating parameters in an update model, calculating an evaluation index of the rock detection model by using a verification set after each round of training is finished, and storing the model weight of the generation with the best evaluation index of the verification set as the final weight after model training is finished after all training rounds are finished. And loading and freezing the optimal model weight during testing, forward reasoning is carried out on the input picture, the position of the rock prediction frame predicted by the model and the confidence coefficient of the category are output, and then the model is visualized into the input feature map after confidence coefficient filtering and NMS non-maximum value suppression.
The method has the highest accuracy and detection rate, is superior to the existing main stream target detection algorithm, and is compared with three main stream target detection methods in the comparison experiment of table 1, and compared with the mAP of the traditional yolov network, the method yolov-MHSA of the invention is improved by 5%, so that the asteroid rock detection task can be completed better.
Table 1 comparison of the detection effects of different algorithms on the celestial body rock
Model P％ R％ F1％ [email protected] [email protected]:.95％ FPS
yolov5 67.3 58.9 62.8 65.2 28.0 76.5
yolov7 71.8 67.5 69.6 69.9 30.1 62.8
yolov8 67.6 65.3 66.4 70.0 33.4 50.2
yolov8-MHSA 74.0 68.3 71.0 75.0 38.6 31.8
While the present invention has been described in detail with reference to the drawings, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the knowledge of those skilled in the art.

Claims (3)

1. A method for detecting rock on the surface of a celestial body, comprising:
S1, constructing a small celestial body surface rock detection reference database;
S2, zooming the pictures in the small celestial body surface rock detection database constructed in the S1 to a preset specification, then sending the pictures into a backbone network to extract all levels of features, and finally enhancing the rock features in the small celestial body feature map through a multi-head self-attention module MHSA in the backbone network, and then obtaining a small celestial body surface picture feature map through SPPF operation;
S3, inputting the small celestial body surface feature map output in the S2 into a Neck network, adding a small target detection layer through the Neck network, fusing high-level semantic features in the low-level feature map, fusing low-level position information and texture features in the high-level feature map, and performing multi-scale fusion on the feature maps of each layer obtained in the S2 to obtain a multi-scale small celestial body surface feature map to be detected;
s4, sending the multiscale celestial body surface feature map to be detected obtained in the S3 into a detection head detect, and predicting the position coordinates and the confidence of the rock in the map;
The Neck network in the S3 adopts a PANet structure, two characteristic transmission channels are arranged in PANet, namely a deep-to-shallow transmission channel and a shallow-to-deep transmission channel, a small target detection layer is added on the PANet structure, the small target detection layer is used for up-sampling the P3 level in the deep-to-shallow transmission channel to the P2 level, the characteristic map is spliced with a characteristic map of the P2 level output by a first C2f module in a backbone network in the channel direction, a first-scale small celestial body surface characteristic map to be detected is output by the first C2f module and used for detecting rock of a first preset size, and the first-scale small celestial body surface characteristic map to be detected is spliced with the output of a second C2f module in the deep-to-shallow transmission channel in the channel direction and then is sent to the first C2f module in the shallow-to-deep transmission channel to output a second-scale small celestial body surface characteristic map to be detected for detecting rock of a second preset size in the PANet transmission channel from the shallow-to deep layer; a second C2f module in the shallow-layer-to-deep-layer transmission channel outputs a third-scale celestial body surface characteristic map to be detected for detecting rocks of a third preset size; and a third C2f module in the shallow-layer-to-deep-layer transmission channel outputs a fourth-scale small celestial body surface characteristic map to be detected for detecting rocks of a fourth preset size.
2. The method according to claim 1, wherein the backbone network in S2 is framed by a yolov backbone network, and MHSA is added before SPPF operation of the backbone network.
3. The method for detecting the surface rock of the celestial body according to claim 1, wherein in the step S4, the multi-scale surface feature map of the celestial body to be detected is sent to a detection head detect for detection, the detection heads share four scales, each detection head is of a decoupling structure, a classification task and a prediction frame position prediction task are respectively completed by different structures, regression parts of the four detection heads respectively predict the positions of prediction frames, each prediction frame parameter comprises a center point x coordinate, a center point y coordinate, a width w and a height h, classification parts of the four detection heads predict n confidence degrees for each prediction frame, and n is the number of prediction categories; and finally mapping the prediction frames to the original image, framing out a target object on the original image, and finally obtaining a detection result.
CN202410085265.XA 2024-01-21 2024-01-21 Small celestial body surface rock detection method Active CN117830813B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202410085265.XA CN117830813B (en) 2024-01-21 2024-01-21 Small celestial body surface rock detection method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202410085265.XA CN117830813B (en) 2024-01-21 2024-01-21 Small celestial body surface rock detection method

Publications (2)

Publication Number Publication Date
CN117830813A CN117830813A (en) 2024-04-05
CN117830813B true CN117830813B (en) 2024-06-11

Family

ID=90517503

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202410085265.XA Active CN117830813B (en) 2024-01-21 2024-01-21 Small celestial body surface rock detection method

Country Status (1)

Country Link
CN (1) CN117830813B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115661622A (en) * 2022-09-27 2023-01-31 北京航空航天大学 Merle crater detection method based on image enhancement and improved YOLOv5
CN117079098A (en) * 2023-07-07 2023-11-17 北京理工大学 Space small target detection method based on position coding
CN117197687A (en) * 2023-03-13 2023-12-08 西南科技大学 Unmanned aerial vehicle aerial photography-oriented detection method for dense small targets
CN117218344A (en) * 2023-09-11 2023-12-12 长春理工大学 Mars surface target segmentation method, device and equipment based on deep neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115661622A (en) * 2022-09-27 2023-01-31 北京航空航天大学 Merle crater detection method based on image enhancement and improved YOLOv5
CN117197687A (en) * 2023-03-13 2023-12-08 西南科技大学 Unmanned aerial vehicle aerial photography-oriented detection method for dense small targets
CN117079098A (en) * 2023-07-07 2023-11-17 北京理工大学 Space small target detection method based on position coding
CN117218344A (en) * 2023-09-11 2023-12-12 长春理工大学 Mars surface target segmentation method, device and equipment based on deep neural network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Tomato Maturity Detection and Counting Model Based on MHSA-YOLOv8;Li, P.; Zheng, J.; Li, P.; Long, H.; Li, M.; Gao, L.;Sensors;20230726;第23卷(第15期);全文 *
UAV-YOLOv8: A Small-Object-Detection Model Based on Improved YOLOv8 for UAV Aerial Photography Scenarios;Wang, G.; Chen, Y.; An, P.; Hong, H.; Hu, J.; Huang, T.;Seonsors;20230815;第23卷(第16期);全文 *
基于多头自注意力机制和PANet的优化YOLOv5行人检测算法;宋子昂, 刘惠临;宁夏师范学院学报;20240115;第45卷(第2024.01期);全文 *

Also Published As

Publication number Publication date
CN117830813A (en) 2024-04-05

Similar Documents

Publication Publication Date Title
CN111091105B (en) Remote sensing image target detection method based on new frame regression loss function
CN112488210A (en) Three-dimensional point cloud automatic classification method based on graph convolution neural network
CN114202696A (en) SAR target detection method and device based on context vision and storage medium
Li et al. Dual-view 3d object recognition and detection via lidar point cloud and camera image
US10943352B2 (en) Object shape regression using wasserstein distance
CN110309842B (en) Object detection method and device based on convolutional neural network
Fan et al. Real‐Time Object Detection for LiDAR Based on LS‐R‐YOLOv4 Neural Network
US12008779B2 (en) Disparity estimation optimization method based on upsampling and exact rematching
CN107247960A (en) Method, object identification method and the automobile of image zooming-out specification area
CN112613350A (en) High-resolution optical remote sensing image airplane target detection method based on deep neural network
CN111507369A (en) Automatic driving vehicle space learning method and device, and testing method and device
CN115661622A (en) Merle crater detection method based on image enhancement and improved YOLOv5
CN113095417B (en) SAR target recognition method based on fusion graph convolution and convolution neural network
Huang et al. Measuring the absolute distance of a front vehicle from an in-car camera based on monocular vision and instance segmentation
Li et al. An aerial image segmentation approach based on enhanced multi-scale convolutional neural network
CN112396039A (en) Mars grid terrain map generation method based on neighborhood relationship
CN115984537A (en) Image processing method and device and related equipment
Gao et al. Road extraction using a dual attention dilated-linknet based on satellite images and floating vehicle trajectory data
Yao et al. Vision-based environment perception and autonomous obstacle avoidance for unmanned underwater vehicle
US10445611B1 (en) Method for detecting pseudo-3D bounding box to be used for military purpose, smart phone or virtual driving based-on CNN capable of converting modes according to conditions of objects and device using the same
CN117830813B (en) Small celestial body surface rock detection method
Chaturvedi et al. Small object detection using retinanet with hybrid anchor box hyper tuning using interface of Bayesian mathematics
CN107170004A (en) To the image matching method of matching matrix in a kind of unmanned vehicle monocular vision positioning
CN116129234A (en) Attention-based 4D millimeter wave radar and vision fusion method
CN112465836B (en) Thermal infrared semantic segmentation unsupervised field self-adaption method based on contour information

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