CN114415254B - Multi-case weak supervision mars surface morphology detection method based on online learning - Google Patents

Abstract

A multi-case weak supervision mars surface morphology detection method based on online learning belongs to the field of machine vision object detection. The method aims at the problems that the existing weak supervision object detection method is low in detection precision and converges on a local optimal solution. Generating a plurality of candidate boxes by adopting a selective search algorithm on the Mars long-shot image; extracting image features of the Mars perspective image by adopting a VGG16 network model; further obtaining the full connection characteristic of each candidate frame; in the weak supervision detection network, inputting the full connection characteristics of the candidate frames, judging the category of each candidate frame through classification and detection branches, scoring the position information of the candidate frames, and finally multiplying the scores of the two branches to obtain the score of the candidate frame as a case level label; and the K-grade refined network layer takes the score of each candidate box of the multi-case learning network or the previous-grade branch as supervision information, trains other optimized branches of the network, and performs backward propagation calculation. The method is used for detecting the surface topography and the target of the mars.

Description

Multi-instance Weakly Supervised Mars Surface Morphology Detection Method Based on Online Learning

technical field

The invention relates to a multi-case weakly supervised Mars surface shape detection method based on online learning, and belongs to the technical field of machine vision object detection.

Background technique

Mars is the terrestrial planet next to Earth in the solar system and the most similar to the Earth in the solar system; the discovery of water has made Mars considered one of the most likely planets to harbor life and a prime target for space exploration one. At present, the aerospace industry has achieved world-renowned achievements in earth satellite and manned spaceflight projects. The development of deep space exploration will be the follow-up focus, which is of great significance to scientific and technological progress and social development. Due to the complex and changeable environment of Mars, which is often accompanied by dust and other weather, the current key technology for Mars exploration is to obtain high-resolution images of Mars, and use the high-resolution images to detect objects/terrains on the surface of Mars. It can land on the surface of Mars smoothly and safely, and then carry out follow-up related scientific research tasks.

Visual tasks are the key link for the Mars rover to achieve autonomous work. At present, the Mars rover is equipped with binocular visual environment perception and autonomous path planning technologies; with the advancement of Mars exploration, more visual technologies will be applied to the Mars rover to achieve more complex detection tasks. The Mars rover is equipped with a Mars rover to carry out planetary exploration missions on Mars. In order to better make the Mars rover land on the surface of Mars smoothly and safely, it is very important for the rover to obtain the topography of the Martian surface.

At present, the classic mainstream frameworks for object detection include YOLO algorithm based on bounding box regression, RCNN series algorithms based on candidate region generation, etc. Carry out case-level labeling (case-level labeling means that when training the network, not only the category of the image is given, but also the target positioning box is given to represent the exact position with the center coordinates and the length and width of the bounding box), through the idea of regression and based on the candidate area The idea of case-level training is carried out, and the category and position of the detected object are further obtained. However, there is no case-level labeling information in the currently collected images of Mars topography and geomorphology. If artificial labeling is used, the subjectivity of case-level labeling will be too strong, and manual labeling is a time-consuming and labor-intensive process. The above-mentioned The condition for instance-level labeling is difficult to satisfy.

In the absence of the above-mentioned case-level annotation information, weakly supervised learning is usually used to achieve object detection tasks. In the object detection task based on weakly supervised learning, only the labeling of object category information (no labeling of position information) is required to achieve the purpose of object detection. However, in the field of weakly supervised object detection, there are two main problems: (1) the model is sensitive to initialization, and the detection accuracy is low; (2) it is easy to converge to the local optimal solution, and the intuitive performance can only detect the most effective The region of the feature, rather than the entire object region, causes object localization to fail.

Contents of the invention

Aiming at the problems that the existing weakly supervised object detection method has low detection accuracy and converges to a local optimal solution, the present invention provides a multi-instance weakly supervised Mars surface morphology detection method based on online learning.

A multi-instance weakly supervised Mars surface form detection method based on online learning of the present invention, comprising:

Use Mars prospect images to form a training set; each Mars perspective image is configured with a terrain category label;

Set up an online network including a candidate frame generation unit, a VGG16 network model and a weakly supervised detection network;

In the candidate frame generation unit, a selective search algorithm is used to generate candidate frames of multiple target objects or target terrains for each Mars prospect image; each Mars prospect image is imaged by using the VGG16 network model pre-trained on ImageNet Feature extraction; combine the image features and the position information of each candidate frame to obtain the fully connected features of each candidate frame;

The weakly supervised detection network is used to detect the fully connected features of each candidate frame, and the position detection score of the candidate frame containing the target object or target terrain is obtained from the detection results; at the same time, the fully connected features of each candidate frame are classified, and classified The result is a category score; the primary case-level label of the candidate box is obtained from the dot product result of the position detection score of the candidate box and the category score;

Using primary case-level labels as supervision information, the fully connected features of each candidate frame are optimized step by step using K-level refined network layers to obtain the final case-level labels of the candidate frames;

Compare the case-level labels output by each level of refined network layer with the case-level labels output by the adjacent first-level refined network layer or the primary case-level labels, and compare the classification results of each candidate box with the ground truth of the terrain category The loss function is obtained by comparison, the weakly supervised detection network is optimized based on the loss function, and the final weakly supervised detection network is obtained;

The final weakly supervised detection network is used to detect the surface morphology of Mars on the Mars perspective images acquired in real time.

According to the online learning-based multi-instance weakly supervised Mars surface morphology detection method of the present invention, the terrain category labels include hills, ravines, rocky areas and flat areas.

According to the online learning-based multi-case weakly supervised Mars surface form detection method of the present invention, the final case-level labeling process of the candidate frame obtained by using a K-level refined network layer includes:

In the first-level refined network layer, the softmax classifier is used to classify the fully connected features of each candidate box at the case level, and the case-level classification results are used to correct the primary case-level labels to obtain the first-level corrected case-level labels;

In the second-level refined network layer, the softmax classifier is used to classify the fully connected features of each candidate box at the case level, and the case-level classification results are used to correct the first-level correction case-level label to obtain the second-level correction case-level label;

...

In the K-level refined network layer, the softmax classifier is used to classify the fully connected features of each candidate box at the case level, and the case-level classification results are used to correct the K-1 level corrected case-level label to obtain the K-level corrected case-level label as the final Instance-level labels.

According to the multi-case weakly supervised Mars surface form detection method based on online learning of the present invention, the VGG16 network model obtains the image features of the Mars prospect image through a convolution operation; then combines the image features to perform an RoI pooling operation on each candidate frame Get the fully connected features of the candidate box.

According to the online learning-based multi-case weakly supervised Mars surface morphology detection method of the present invention, each Mars prospect image generates 1000 candidate frames of target objects or target terrains.

According to the multi-instance weakly supervised Mars surface form detection method based on online learning of the present invention, the VGG16 network model is backpropagated through weights with 16 layers of parameters.

According to the multi-instance weakly supervised Mars surface morphology detection method based on online learning of the present invention, the Mars perspective image is provided by the mars32k data set and the GMSRI data set of NASA.

According to the online learning-based multi-case weakly supervised Mars surface morphology detection method of the present invention, after the online network is trained, the online network is also verified using a verification set, and a test set is used for testing; training set, verification set and test set The ratio is 3:1:1.

Beneficial effects of the present invention: the present invention can be used for detection of Martian surface objects and topography in the case of scarcity of training data labeling information in Mars exploration-related tasks. It is used to perform the detection task of the terrain during the actual exploration of the Mars rover when it lands, so that the Mars rover can better find the most suitable position for the rover to land.

The method of the present invention completes the detection of the topography of Mars only through image-level annotation information, providing an opportunity for the probe to land on Mars and select the most scientifically valuable planetary surface position for subsequent planning tasks. Among them, in the multi-instance weakly supervised network, the input is the fully connected feature of each candidate box, and the functions of the two parallel classification and detection branches are to judge the category of each candidate box and score the position information of each candidate box , and finally multiply the scores of the classification branch and the detection branch to obtain the score of the candidate box as the case-level label of the candidate box. In the K-level refined network layer, the multi-instance learning network or the score of each candidate frame of the previous branch is used as supervision information to train other optimization branches of the network, and perform backward propagation calculations to further improve detection accuracy.

The method of the present invention breaks through the limitation that the full-supervised object detection method requires case-level labeling information, solves the problems of low detection accuracy and convergence to a local optimal solution in the weakly supervised object detection method, and does not need to manually mark case-level labels. Utilizing the method of the invention, the purpose of detecting Martian surface objects and topography can be achieved under the supervision information of only the image level (that is, the topography and landform type existing in each image).

As the basic technical research work in the Mars exploration mission, the method of the invention can promote the follow-up planning task to a certain extent, and provide certain technical support for the follow-up planning task.

Description of drawings

Fig. 1 is the schematic diagram of the principle of the multi-instance weak supervision Mars surface form detection method based on online learning according to the present invention;

Figure 2 is a schematic diagram of the comparison of Mars vision images of four terrain categories;

Fig. 3 is a schematic diagram of packages and instances of the Mars vision image;

Figure 4 is a schematic diagram of the structure of the VGG16 network model;

FIG. 5 is a schematic diagram of the principle of a K-level refined network layer;

Fig. 6 is a comparison chart of experimental results of specific embodiments.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

It should be noted that, in the case of no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but not as a limitation of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS 1. As shown in FIGS. 1 to 5, the present invention provides a multi-instance weakly supervised Mars surface morphology detection method based on online learning, including:

Use Mars prospect images to form a training set; each Mars perspective image is configured with a terrain category label;

Set up an online network including a candidate frame generation unit, a VGG16 network model and a weakly supervised detection network;

In the candidate frame generation unit, a selective search algorithm is used to generate candidate frames of multiple target objects or target terrains for each Mars prospect image; each Mars prospect image is imaged by using the VGG16 network model pre-trained on ImageNet Feature extraction; combine the image features and the position information of each candidate frame to obtain the fully connected features of each candidate frame;

The weakly supervised detection network is used to detect the fully connected features of each candidate frame, and the position detection score of the candidate frame containing the target object or target terrain is obtained from the detection results; at the same time, the fully connected features of each candidate frame are classified, and classified The result is a category score; the primary case-level label of the candidate box is obtained from the dot product result of the position detection score of the candidate box and the category score;

Using primary case-level labels as supervision information, the fully connected features of each candidate frame are optimized step by step using K-level refined network layers to obtain the final case-level labels of the candidate frames;

Compare the case-level labels output by each level of refined network layer with the case-level labels output by the adjacent first-level refined network layer or the primary case-level labels, and compare the classification results of each candidate box with the ground truth of the terrain category The loss function is obtained by comparison, the weakly supervised detection network is optimized based on the loss function, and the final weakly supervised detection network is obtained;

The final weakly supervised detection network is used to detect the surface morphology of Mars on the Mars perspective images acquired in real time.

This embodiment can realize the detection of Martian surface objects and terrain in Mars images, especially so that the Martian surface object and terrain detection technology no longer depends on a large database with label information, and only needs to know the simple object or terrain category information in the image. The labeling of complex object position information is required to train the model, and then achieve the purpose of detecting objects and terrain on the surface of Mars. It mainly realizes the detection of Martian topography through the feature recognition of Martian surface objects (such as stones and fine sand), and creates conditions for the Mars rover landing or the selection of the most scientifically valuable landing site in the Mars rover’s subsequent exploration missions.

As an example, as shown in FIG. 2 , the terrain category labels include hills, ravines, stone areas, and flat areas. These four category labels will be used as image-level labels for each Mars image to train the online network.

Further, as shown in Figures 1 and 5, the final instance-level labeling process of the candidate box obtained by using the K-level refined network layer includes:

In the first-level refined network layer, the softmax classifier is used to classify the fully connected features of each candidate box at the case level, and the case-level classification results are used to correct the primary case-level labels to obtain the first-level corrected case-level labels;

In the second-level refined network layer, the softmax classifier is used to classify the fully connected features of each candidate box at the case level, and the case-level classification results are used to correct the first-level correction case-level label to obtain the second-level correction case-level label;

...

In the K-level refined network layer, the softmax classifier is used to classify the fully connected features of each candidate box at the case level, and the case-level classification results are used to correct the K-1 level corrected case-level label to obtain the K-level corrected case-level label as the final Instance-level labels.

The present embodiment utilizes a multi-instance learning method to implement a weakly supervised object and terrain detector. Since there is no human-labeled position information as the true value of object position regression, weakly supervised object and terrain detectors usually converge to a local optimal solution, leading to local focusing phenomena, such as only highlighting the local most discriminative of objects and terrain area of force. In order to improve the recognition rate, the present invention proposes an online optimization strategy, which includes several optimization branches parallel to the multi-instance learning and detection network, as shown in FIG. 1 . During training, the weakly supervised detection network or the output result of the previous refined network layer is used as supervision information to supervise and train the optimization branch of the subsequent refined network layer, and further achieve the purpose of accurate detection of Martian surface objects and terrain.

In summary, the method of the present invention includes an object detection part (weakly supervised detection network based on multiple instances) and an optimization part (K-level refined network layer). Then perform element-wise operation to obtain a C×R matrix, where C is the type of target, that is, rocky areas, ravines, hilly areas, and flat areas, and R is the number of candidate frames. Then, through the K-th refinement network, the local focus of the target frame is reduced, so that the output position of the object detection is more accurate and meets the actual needs.

The K-level refined network layer is not alternately trained, the trained data is re-labeled, and the idea of further iterating multiple trainings is repeated, but the area with the highest score candidate area whose intersection and union ratio is greater than a certain threshold The homogeneous method is repeatedly refined, and the case-level labels of each candidate area are further corrected, so that a multi-instance weakly supervised Martian surface object and terrain detection network based on online learning can be obtained, and then the existing multi-instance weak supervision network can be solved. The position of the supervisory detector is not accurate and the detection accuracy is low.

Further, as shown in Figures 1 and 4, the VGG16 network model obtains the image features of the Mars vision image through the convolution operation; then combines the image features for each candidate frame through the RoI pooling operation to obtain the full connection of the candidate frame feature.

As an example, as shown in FIG. 1 , each Mars distant image generates 1000 candidate frames of target objects or target terrains.

In the weakly supervised detection network, for each piece of Mars image data that only has the label of the Martian terrain and landform category, a selective search algorithm is first used to generate about 1,000 possible locations of target objects and regions, and the possible locations of each target object and region The position is called the candidate area (proposals), marked with the candidate box; then use the VGG16 network model to obtain the fully connected features of each candidate area, and these extracted fully connected features are input into the subsequent detection and classification network to realize the Mars image. object and terrain detection tasks.

The VGG16 network model is backpropagated through weights with 16 layers of parameters.

As an example, the Mars distant image is provided by NASA's mars32k dataset and the GMSRI dataset.

In actual use of this embodiment, the category of the specific detection object may be determined according to the actual problem of the user.

In this embodiment, during the training process, the real Mars images taken by the Curiosity rover and the Mars images generated by the learning method can be used to form a data set. The dataset provided by NASA contains 32,368 color images collected by the Curiosity rover on Mars from August 2012 to November 2018 and about 30,000 color Martian images generated. These images show various geographical and geological features of Mars, such as mountains and valleys, craters, dunes and rocky terrain, etc. The image resolution is 560×500px. The mars32k and GMSRI data set images include close-up views of Mars, long-range views, larger rocks, fine sand grains, and some images taken in dim light. Since the probe needs to search for flat areas when it is kilometers away from the surface of Mars, it uses Mars prospect images are used as a training set, and the image categories are divided into four categories: hills, ravines, rocky areas, and flat areas. Using these only category labels to train the online network does not require labeling of location information, and does not require a lot of manpower and material resources to label the database, which can avoid the bias introduced by the subjectivity of human labeling.

Furthermore, after the online network is trained, the online network is also verified with a verification set and tested with a test set; the ratio of the training set, verification set, and test set is 3:1:1.

To sum up, this embodiment can obtain a higher detection rate of surface objects and terrains on Mars without the need for location labeling information. It can promote the development of object and terrain detection technology based on Mars images, and provide certain technical support for subsequent tasks related to Mars exploration.

Specific examples:

Firstly, training samples are prepared according to the actual needs of users, and then a weakly supervised object/terrain detector is trained according to the multiple instance learning (MIL) method. After that, the online optimization strategy is used to further improve the detection accuracy of weakly supervised objects and terrain, and obtain more accurate detection results of Martian surface terrain. Each part is described in detail below:

First prepare the training samples.

A weakly supervised detection network based on multi-instance learning is then trained.

The weakly supervised detection network is an end-to-end detection method, wherein in multi-instance learning, the label of the package is known, and the label of the case is unknown, and the label of the package only shows what kind of target exists in this pair of images , but the location of the target is unknown. Taking the Mars image as an example, each 560×500 Mars image is a package, and some areas (Patches) in the image are examples, as shown in Figure 3. It can be seen from Figure 3 that there are hills and rocks in the package, but no ravines; there are only category labels for the package:

y=[y ₁ ,y ₂ ,y ₃ ,...,y _n ]∈RC ^×l ,

Among them, y _i indicates that the target is a label of type i, i=1, 2, 3,..., n; n is the number of regions (Patches) in the image, and C is the number of category labels;

The value of y _i is +1 or -1.

If the labeled category label is {1,-1,1}, it means that there are examples of hilly areas, no examples of gully areas, and examples of rocky areas, and there are no examples of hilly areas and rocky areas in the labels of training samples. specific location information. Next, for each input image of Mars, a selective search algorithm is used to generate 1000 possible locations of target objects/terrains, and each possible location of target objects/terrains is called a candidate area (proposals). Then, use the pre-trained VGG16 network model to extract features, and finally use the RoI pooling method to obtain the features of each candidate area, and then obtain the fully connected features of each candidate area, and these extracted fully connected features are input to the following In the detection and classification network, the task of object/terrain detection in Mars images is realized. Among them, the VGG16 network structure is named after backpropagation with weights of 16 layers of parameters. The network structure includes the following parts: feature extraction network, fully connected network, and two parallel branch detection and recognition networks, as shown in Figure 4 Architecture for Weakly Supervised Martian Surface Object/Terrain Detector Networks Based on Multiple Instance Learning.

Design online optimization strategies. In order to prevent the output result of the bounding box from easily appearing in the object area with strong local information, which is not an accurate and complete position, this embodiment proposes a K-time refinement correction strategy for online learning on the basis of a weakly supervised detector, aiming at Realize the ability to transfer label information from a small area of a high-scoring bounding box to a large area, as shown in Figure 5. Since there are only image-level labels and no case-level labels in weakly supervised detection, the idea of the present invention is to calculate the cross-entropy loss function through the case set labels and the score matrix, and update the score matrix through backpropagation, so the case-level labels are obtained. important. First, the {C+1}-dimensional candidate area is the probability of each class, and the second output matrix is an R×C matrix. The loss backpropagation is optimized for this matrix through the case-level labels output by the previous refinement. Score matrix, each candidate area will generate a C+1-dimensional vector, k represents the current optimization times, where k={0,1...,K-1}.

NASA's mars32k and GMSRI data sets are used as the training, verification and verification data of the invention. 60% of the data set is used as the training set, and the loss is reduced through the number of iterations. 20% of the data set is used as the verification set, which is verified after a certain number of iterations. At this time, the generalization ability of the model, the remaining 20% of the data set is used as the test set, which proves the generalization ability of the network structure and the effectiveness of the algorithm in the detection of Mars topography. Select the pre-trained VGG16 to extract the features of the Martian topography, and use 20 epoch iterations when training the weakly supervised basic detector. The learning rate of the first 10 epochs is 10 ^-5 , and the learning rate of the last 10 epochs is 10 ^-6 . The optimizer chooses Adam, and this optimizer introduces the momentum method, which makes the parameter update deviate from the local optimum, and the momentum and weight decay are set to 0.9 and 0.0005, respectively. After the online optimization strategy is introduced, the network training adopts the optimizer of stochastic gradient descent, and the mini-batch of each optimized input image is 2 (mini-batch=2), the momentum and weight decay are still set to 0.9 and 0.0005, and the refinement When the Iou threshold is It=0.5. Finally, the non-maximum suppression NMS is set to 0.3 to calculate the final evaluation indicators mAP and CorLoc. The number of refinements is selected as K=4 through experiments. In order to increase the training data samples, the method of data augmentation is adopted, the probability of random horizontal flipping of the input Mars image is set to p=0.5, and the ratio of the length and width of the image is fixed between 0.5-2. If the above conditions are not met, the training set, validation set, test set set the shortest side to {480, 576, 688, 864, 1200}, and the longest side does not exceed 2000 pixels.

The multi-instance weakly supervised Martian surface object and terrain detection network based on online learning trained by the above steps can realize the Martian surface object and terrain detection function without the need for instance-level labeling information, and can be applied to Mars rover landing according to specific needs The selection of points, as well as the selection of the most scientifically valuable landing points in the follow-up planning tasks of Mars exploration, are not limited by the existing object detection methods that require complete information labeling databases, and do not need to spend manpower and material resources on each training sample. Annotate the image.

The experiment proves that the method of the present invention has high detection accuracy and accurate positioning for the surface terrain of Mars. Table 1 and Standard 2 are comparative data of experimental results:

Table 1 Comparison data table of experimental results with mAP (%) as evaluation index

Table 2 Take CorLoc (%) as the experimental result comparison data table of the evaluation index

Among them, mAP is the mean average precision, which is an index for evaluating the test samples, and Corloc is the correct location rate (Correct Location), which is an index for evaluating the positioning effect of the training samples during the training process. It can be seen from the comparative data that only weak supervision detection mAP/Corloc is 50.6%/69.4%, compared with the introduction of online optimization strategy to increase mAP/Corloc to 58.3%/78.3, which significantly improves the local area of weak supervision Focusing on the problem, the effectiveness of the online optimization strategy is demonstrated. Furthermore, the fully supervised detector that uses the generated false ground truth as the training case-level label fully utilizes the strong regression ability of fully supervised learning. Compared with the output of the weakly supervised object detector, the final position coordinate information of the object is more accurate. For accuracy, the effectiveness of the method of the present invention is further proved from the side. Figure 6 is a diagram of the experimental results. During the test, the position of the topography on the surface of Mars is detected, and the candidate bounding box selects the objects and topography on the surface of Mars; the rock area, hill area, The ravine area further completes the selection of flat areas in the whereabouts task. Large rocks, dense rock areas, hill areas, and ravine areas can be clearly observed from the detection area.

Although the invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It is therefore to be understood that numerous modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the invention as defined by the appended claims. It shall be understood that different dependent claims and features described herein may be combined in a different way than that described in the original claims. It will also be appreciated that features described in connection with individual embodiments can be used in other described embodiments.

Claims (4)

1. A multi-case weak supervision mars surface morphology detection method based on online learning is characterized by comprising the following steps,

forming a training set by using a Mars perspective image; configuring a terrain category label for each Mars long shot image;

the online network is arranged to comprise a candidate box generation unit, a VGG16 network model and a weak supervision detection network;

a candidate frame generating unit generates a plurality of candidate frames of target objects or target terrains for each Mars long shot image by adopting a selective search algorithm; extracting image features of each Mars perspective image by adopting a VGG16 network model pre-trained on ImageNet; combining the image characteristics and the position information of each candidate frame to obtain the full-connection characteristics of each candidate frame;

detecting the full-connection characteristics of each candidate frame by adopting a weak supervision detection network, and obtaining a position detection score of the candidate frame including a target object or a target terrain according to a detection result; classifying the full-connection characteristics of each candidate frame, and obtaining a classification score according to a classification result; obtaining a primary case-level label of the candidate box according to the dot product result of the position detection score and the category score of the candidate box;

taking the primary case-level label as supervision information, and performing optimization processing on the full connection characteristics of each candidate frame step by adopting a K-grade fine network layer to obtain a final case-level label of the candidate frame;

comparing the case-level label output by each refinement network layer with the case-level label or the primary case-level label output by the adjacent previous refinement network layer, comparing the classification result of each candidate box with the terrain category true value to obtain a loss function, and optimizing the weak supervision detection network based on the loss function to obtain a final weak supervision detection network;

carrying out Mars surface morphology detection on the Mars long shot image acquired in real time by adopting a final weak supervision detection network;

the process of obtaining the final case level label of the candidate box by adopting the K-level refinement network layer comprises the following steps:

in a first-level refinement network layer, performing case-level classification on the full-connection features of each candidate frame by adopting a softmax classifier, and correcting the primary case-level labels by adopting case-level classification results to obtain first-level corrected case-level labels;

in a second-level refinement network layer, performing case-level classification on the full-connection features of each candidate frame by adopting a softmax classifier, and correcting a first-level correction case-level label by adopting a case-level classification result to obtain a second-level correction case-level label;

……

in a K-level refinement network layer, performing case-level classification on the full-connection features of each candidate frame by adopting a softmax classifier, and correcting a K-1-level correction case-level label by adopting a case-level classification result to obtain a K-level correction case-level label as a final case-level label;

generating 1000 candidate frames of target objects or target terrains by each Mars long shot image;

the VGG16 network model performs back propagation through weights with 16-layer parameters;

the Mars perspective image is provided by the mars32k dataset and the GMRI dataset of NASA;

the VGG16 network model adopts 20 epoch iteration times when training the weak supervision detection network, and the learning rate of the first 10 epochs is 10 ^-5 And the learning rate of the last 10 epochs is 10 ^-6 (ii) a The optimizer selects Adam.

2. The method of claim 1, wherein the terrain category labels comprise hills, ravines, stone regions, and flat regions.

3. The multi-case weak supervision mars surface morphology detection method based on online learning of claim 2, wherein the VGG16 network model obtains image features of mars perspective images through convolution operation; and combining the image characteristics to obtain the full connection characteristics of the candidate frames for each candidate frame through the RoI pooling operation.

4. The multi-case weak supervision mars surface morphology detection method based on online learning of claim 3, wherein after the online network is trained, the online network is verified by using a verification set, and is tested by using a test set; the proportion of the training set, the verification set and the test set is 3:1:1.

Images