Deep space semantic segmentation pseudo label learning method and system based on error correction output code

Abstract

The application discloses a deep space semantic segmentation pseudo label learning method and system based on error correction output code and relates to the technical field of computer vision.The application comprises the following steps: receiving a label-free image of a planetary surface environment; constructing a teacher model-student model collaborative supervision framework, wherein the student model performs feature learning on the label-free image processed by strong enhancement and outputs K-dimensional attribute bit prediction; the teacher model dynamically inherits the student model parameters by exponential moving average, extracts stable representation from the label-free image processed by weak enhancement, and generates bit-level pseudo labels; and the bit-level pseudo labels generated by the teacher model are denoised.The application introduces an error correction output code to construct a fine-grained category coding mechanism, solves the problem of pseudo label error amplification in deep space semantic segmentation, improves the segmentation accuracy and robustness in a few-label scenario, and provides reliable environment perception information for autonomous decision-making of a probe.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a pseudo-label learning method and system for deep space semantic segmentation based on error-correcting output codes. Background Technology

[0002] In deep space exploration missions, semantic segmentation of planetary surface environments is a core component for autonomous decision-making. Accurate identification of various targets on the planetary surface, including rocks, soil, craters, and equipment components, is crucial. This information forms the basis for critical operations such as path planning, obstacle avoidance, and sample collection. However, the unique characteristics of deep space exploration present significant challenges to semantic segmentation. The most prominent challenge is the scarcity of labeled data. Unknown planetary surfaces lack labeled sample data, making it difficult to train high-performance segmentation models using traditional fully supervised learning methods. Therefore, pseudo-label learning paradigms such as unsupervised domain adaptation and semi-supervised learning become essential. These methods can utilize limited labeled data or cross-domain data to assist model training, playing a vital role in scenarios with limited labels.

[0003] While pseudo-label learning is widely used in semantic segmentation, existing methods have significant limitations in the unique scenario of deep space exploration. First, planetary surfaces contain numerous similar targets; different types of rocks, and small craters and surface depressions, often share highly similar visual features, making it easy to generate incorrect pseudo-labels during the learning process. Existing methods generally employ one-hot encoding to process pseudo-labels, which amplifies local prediction errors, causing them to accumulate during training and ultimately leading to misjudgments of the environment by the probe, affecting the reliability of subsequent decisions. Second, existing methods rely heavily on high-confidence pseudo-label selection or confidence-based weight allocation strategies when dealing with pseudo-label noise. High-confidence selection strategies only retain pseudo-labels with high model prediction confidence for training, leading the model to overemphasize easily identifiable samples while ignoring those that are difficult to distinguish. While confidence-based weight allocation strategies attempt to balance training by assigning different weights to different pseudo-labels, this method requires carefully designed weight functions and hyperparameters, making it difficult to adapt to unknown scenarios. Third, existing methods fail to fully utilize the shared properties between targets on planetary surfaces. In reality, many seemingly similar targets share common visual or physical properties. For example, rocks and craters are both hard surfaces, while soil and certain sedimentary layers are both soft. These shared properties could provide models with additional distinguishing criteria, helping to alleviate confusion between similar categories. However, existing methods do not consider this in their design, resulting in models lacking effective distinguishing cues when faced with similar targets. This further affects the accuracy of semantic segmentation and makes it difficult to meet the high reliability requirements of environmental perception for deep space exploration missions. Summary of the Invention

[0004] The purpose of this invention is to provide a pseudo-label learning method and system for deep space semantic segmentation based on error correction output codes, aiming to solve the problem of pseudo-label noise caused by label scarcity and high target similarity in the deep space environment.

[0005] According to a first aspect of the present invention, in order to achieve the above-mentioned objective, the present invention provides the following technical solution: a pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes, comprising the following steps:

[0006] Receive unlabeled images of the planetary surface environment;

[0007] A collaborative supervision framework of teacher model and student model is constructed, in which the student model learns features from unlabeled images that have undergone strong enhancement and outputs K-dimensional attribute bit predictions; the teacher model dynamically inherits the parameters of the student model through exponential moving average, extracts stable representations from unlabeled images that have undergone weak enhancement, and generates bit-level pseudo-labels.

[0008] The bit-level pseudo-labels generated by the teacher model are denoised to obtain denoised hybrid pseudo-labels, in which the student model uses the hybrid pseudo-labels as supervision signals to optimize prediction.

[0009] The student model is trained using the overall loss function until the overall loss function is minimized. The trained student model is then used to achieve high-precision semantic segmentation of planetary surface environment images.

[0010] Furthermore, the teacher model dynamically inherits the student model parameters through exponential moving averages, extracts stable representations from the weakly enhanced unlabeled images, and generates bit-level pseudo-labels, as follows:

[0011] (21) Construct an N×K binary codebook matrix Where N is the number of deep space target categories, K is the code length, and each category corresponds to a codeword of length K. The codewords are generated through max-min distance encoding or text base encoding to enhance the distinguishability of similar targets.

[0012] (22) Replace the traditional N-class classifier with K binary classifiers, and each binary classifier predicts the probability of the attribute bit using the sigmoid function:

[0013]

[0014] in For pixel features, The weights of the k-th classifier;

[0015] (23) Determine the final category using a soft Hamming distance nearest neighbor query:

[0016]

[0017] The soft Hamming distance is:

[0018]

[0019] in It is the codebook of the nth class.

[0020] Furthermore, the bit-level pseudo-labels generated by the teacher model are denoised to obtain denoised hybrid pseudo-labels. The student model uses the hybrid pseudo-labels as supervision signals to optimize prediction, as detailed below:

[0021] (31) Generate bit-level pseudo-labels and obtain binary codes by applying threshold truncation to the outputs of K binary classifiers;

[0022] (32) Generate codebook-level pseudo-tags and correct error bits by matching the closest codeword from the binary codebook;

[0023] (33) Based on the reliable bit mining algorithm, the shared bits of the candidate category set are extracted as reliable bits, and finally fused to generate a hybrid pseudo-label:

[0024]

[0025] in For reliable bitmasks, These are codebook-level pseudo-tags. These are bit-level pseudo-tags.

[0026] Furthermore, the student model is trained using an overall loss function, which is as follows:

[0027] (41) Binary cross-entropy loss, independently optimizing the classification of each attribute bit:

[0028]

[0029] (42) Pixel-codeword distance loss enhances intra-class compactness by minimizing the cosine distance between the prediction probability and the codeword:

[0030]

[0031] (43) Pixel-codeword contrast loss, used to enhance inter-class separation:

[0032]

[0033] (44) The total loss function is:

[0034]

[0035] in and To balance the weights.

[0036] According to a second aspect of the present invention, the present invention provides a pseudo-label learning system for deep space semantic segmentation based on error-correcting output codes, used to implement the pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes described in the first aspect, comprising:

[0037] A receiving module for receiving unlabeled images of the planetary surface environment;

[0038] The model building module is used to construct a teacher-student model collaborative supervision framework. The student model learns features from unlabeled images that have undergone strong enhancement and outputs K-dimensional attribute bit predictions. The teacher model dynamically inherits the parameters of the student model through exponential moving average and extracts stable representations from unlabeled images that have undergone weak enhancement to generate bit-level pseudo-labels.

[0039] The denoising module is used to denoise the bit-level pseudo-labels generated by the teacher model to obtain denoised hybrid pseudo-labels, in which the student model uses the hybrid pseudo-labels as supervision signals to optimize prediction.

[0040] The training module is used to train the student model using the overall loss function until the overall loss function is minimized to obtain the trained student model. The trained student model is then used to achieve high-precision semantic segmentation of planetary surface environment images.

[0041] According to a third aspect of the present invention, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein when the processor loads and executes the computer program, it employs the deep space semantic segmentation pseudo-label learning method based on error correction output codes described in the first aspect.

[0042] According to a third aspect of the present invention, the present invention provides a storage medium containing computer-executable instructions, wherein the computer-executable instructions, when executed by a computer processor, are used to perform the deep space semantic segmentation pseudo-label learning method based on error-correcting output codes described in the first aspect.

[0043] The present invention has at least the following beneficial effects:

[0044] 1. This invention solves the problem of pseudo-label error amplification in deep space semantic segmentation by introducing error correction output codes to construct a fine-grained category coding mechanism, improves the segmentation accuracy and robustness in scenarios with few labels, and provides reliable environmental perception information for the detector's autonomous decision-making.

[0045] 2. The mid-level label denoising mechanism of this invention generates higher quality pseudo-labels, providing sufficient and robust supervision for unlabeled images, and enhances the intra-class compactness and inter-class separability of features, further improving model performance.

[0046] 3. This invention is easy to integrate with existing methods. It can achieve significant performance improvements based on different segmentation architectures in multiple unsupervised domain adaptation and semi-supervised learning benchmark tests, and has strong applicability and scalability.

[0047] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method described in this invention;

[0049] Figure 2 This is a schematic diagram illustrating the framework of the method described in this invention. Detailed Implementation

[0050] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0051] Example 1:

[0052] Please see Figures 1-2 This invention provides a technical solution: a pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes, comprising:

[0053] S1. Receive unlabeled images of the planetary surface environment;

[0054] S2. Construct a teacher-student model collaborative supervision framework, in which the student model learns features from the strongly enhanced unlabeled image and outputs K-dimensional attribute bit predictions;

[0055] The teacher model dynamically inherits the parameters of the student model through exponential moving averages, extracts stable representations from unlabeled images that have undergone weak enhancement processing, and generates bit-level pseudo-labels, as detailed below:

[0056] (S21) Construct an N×K binary codebook matrix Where N is the number of deep space target categories and K is the code length. Each category corresponds to a codeword of length K. The codewords are generated by max-min distance encoding (maximizing the Hamming distance between categories) or text base encoding (based on the semantic attributes of the target) to enhance the distinguishability of similar targets.

[0057] (S22) Replace the traditional N-class classifier with K binary classifiers. Each binary classifier predicts the probability of the attribute bit using the sigmoid function:

[0058]

[0059] in For pixel features, The weights of the k-th classifier;

[0060] (S23) Determine the final category using a soft Hamming distance nearest neighbor query:

[0061]

[0062] The soft Hamming distance is:

[0063]

[0064] in It is the codebook of the nth class;

[0065] Even if some attribute bits are predicted incorrectly in this way, such as misjudging the "edge" attribute of rocks under low light, classification stability can still be maintained by sharing attributes such as "hardness".

[0066] S3. Denoise the bit-level pseudo-labels generated by the teacher model to obtain denoised hybrid pseudo-labels. The student model uses the hybrid pseudo-labels as supervision signals to optimize prediction, as detailed below:

[0067] (S31) Generate bit-level pseudo-labels and obtain binary codes by applying threshold truncation to the outputs of K binary classifiers;

[0068] (S32) Generate codebook-level pseudo-tags and correct error bits by matching the closest codeword from the binary codebook;

[0069] (S33) Based on the reliable bit mining algorithm, the shared bits of the candidate category set are extracted as reliable bits, and finally fused to generate a hybrid pseudo-label:

[0070]

[0071] in For reliable bitmasks, These are codebook-level pseudo-tags. These are bit-level pseudo-tags;

[0072] This is used to process pseudo-label noise in unlabeled deep space images. For example, changes in lighting may cause rocks to be confused with soil. The reliability is improved by fusing the two pseudo-label forms.

[0073] S4. Train the student model using the overall loss function until the overall loss function is minimized to obtain the trained student model. Use the trained student model to achieve high-precision semantic segmentation of planetary surface environment images, as detailed below:

[0074] Based on binary cross-entropy loss:

[0075]

[0076] The classification of each attribute bit is optimized independently; at the same time, pixel-codeword distance loss is introduced:

[0077]

[0078] Intra-class compactness is enhanced by minimizing the cosine distance between predicted probabilities and codewords, such as feature aggregation of rocks of different sizes; and pixel-codeword contrastive loss is also employed.

[0079]

[0080] Strengthen inter-class separation, such as distinguishing the characteristics of rocks and soil;

[0081] The total loss function is:

[0082]

[0083] in and To balance the weights and effectively improve the model's adaptability to the complex environment of deep space;

[0084] By collaborating with multiple loss functions, the intra-class compactness and inter-class separability of deep space target features are enhanced.

[0085] In summary, the error-correcting output code-based classifier of this invention can decompose categories into attributes, handle some inaccurate bits, and improve the stability and generalization of pseudo-label learning. Secondly, the bit-level label denoising mechanism generates higher-quality pseudo-labels, providing sufficient and robust supervision for unlabeled images. Furthermore, the customized optimization criteria enhance the intra-class compactness and inter-class separability of features, further improving model performance. In addition, this method is easy to integrate with existing methods, and it can achieve significant performance improvements based on different segmentation architectures in multiple unsupervised domain adaptation and semi-supervised learning benchmark tests, demonstrating strong applicability and scalability.

[0086] Example 2:

[0087] This embodiment provides a pseudo-label learning system for deep space semantic segmentation based on error-correcting output codes, used to implement the pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes described in Embodiment 1, including:

[0088] A receiving module for receiving unlabeled images of the planetary surface environment;

[0089] The model building module is used to construct a teacher-student model collaborative supervision framework. The student model learns features from unlabeled images that have undergone strong enhancement and outputs K-dimensional attribute bit predictions. The teacher model dynamically inherits the parameters of the student model through exponential moving average and extracts stable representations from unlabeled images that have undergone weak enhancement to generate bit-level pseudo-labels.

[0090] The denoising module is used to denoise the bit-level pseudo-labels generated by the teacher model to obtain denoised hybrid pseudo-labels, in which the student model uses the hybrid pseudo-labels as supervision signals to optimize prediction.

[0091] The training module is used to train the student model using the overall loss function until the overall loss function is minimized to obtain the trained student model. The trained student model is then used to achieve high-precision semantic segmentation of planetary surface environment images.

[0092] Example 3:

[0093] The present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it adopts the deep space semantic segmentation pseudo-label learning method based on error correction output code described in Embodiment 1.

[0094] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server, and the terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0095] Furthermore, the processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.

[0096] Example 4:

[0097] The present invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the deep space semantic segmentation pseudo-label learning method based on error-correcting output codes described in Embodiment 1.

[0098] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on," "mounted on," "fixed to," or "set on" another element, it may be directly on the other element or there may be an intermediate element present. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be an intermediate element present. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible embodiments.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0102] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. A pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes, characterized in that, Includes the following steps: Receive unlabeled images of the planetary surface environment; A collaborative supervision framework of teacher model and student model is constructed, in which the student model learns features from unlabeled images that have undergone strong enhancement and outputs K-dimensional attribute bit predictions; the teacher model dynamically inherits the parameters of the student model through exponential moving average, extracts stable representations from unlabeled images that have undergone weak enhancement, and generates bit-level pseudo-labels. The bit-level pseudo-labels generated by the teacher model are denoised to obtain denoised hybrid pseudo-labels, in which the student model uses the hybrid pseudo-labels as supervision signals to optimize prediction. The student model is trained using the overall loss function until the overall loss function is minimized to obtain the trained student model. The trained student model is then used to achieve high-precision semantic segmentation of planetary surface environment images. The teacher model dynamically inherits the parameters of the student model through exponential moving averages, extracts stable representations from unlabeled images that have undergone weak enhancement processing, and generates bit-level pseudo-labels, as detailed below: Construct an N×K binary codebook matrix Where N is the number of deep space target categories and K is the code length. Each category corresponds to a codeword of length K. The codewords are generated through max-min distance encoding or text base encoding to enhance the distinguishability of similar targets. Instead of the traditional N-class classifier, K binary classifiers are used, each of which predicts the probability of an attribute bit using a sigmoid function: in For pixel features, The weights of the k-th classifier; The final category is determined using a soft Hamming distance nearest neighbor query: The soft Hamming distance is: in It is the codebook of the nth class.

2. The pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes according to claim 1, characterized in that: The bit-level pseudo-labels generated by the teacher model are denoised to obtain denoised hybrid pseudo-labels. The student model uses the hybrid pseudo-labels as supervision signals to optimize prediction, as detailed below: (21) Generate bit-level pseudo-labels and obtain binary codes by applying threshold truncation to the outputs of K binary classifiers; (22) Generate codebook-level pseudo-tags and correct error bits by matching the closest codeword from the binary codebook; (23) Based on the reliable bit mining algorithm, the shared bits of the candidate category set are extracted as reliable bits, and finally fused to generate a hybrid pseudo-label: in For reliable bitmasks, These are codebook-level pseudo-tags. These are bit-level pseudo-tags.

3. The pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes according to claim 1, characterized in that: The student model is trained using an overall loss function, which is as follows: (31) Binary cross-entropy loss, independently optimizing the classification of each attribute bit: (32) Pixel-codeword distance loss enhances intra-class compactness by minimizing the cosine distance between the prediction probability and the codeword: (33) Pixel-codeword contrast loss, used to enhance inter-class separation: (34) The total loss function is: in and To balance the weights.

4. A pseudo-label learning system for deep space semantic segmentation based on error-correcting output codes, used to implement the pseudo-label learning method for deep space semantic segmentation based on error-correcting output codes as described in any one of claims 1 to 3, characterized in that, include: A receiving module for receiving unlabeled images of the planetary surface environment; The model building module is used to construct a teacher-student model collaborative supervision framework. The student model learns features from unlabeled images that have undergone strong enhancement and outputs K-dimensional attribute bit predictions. The teacher model dynamically inherits the parameters of the student model through exponential moving average and extracts stable representations from unlabeled images that have undergone weak enhancement to generate bit-level pseudo-labels. The denoising module is used to denoise the bit-level pseudo-labels generated by the teacher model to obtain denoised hybrid pseudo-labels, in which the student model uses the hybrid pseudo-labels as supervision signals to optimize prediction. The training module is used to train the student model using the overall loss function until the overall loss function is minimized to obtain the trained student model. The trained student model is then used to achieve high-precision semantic segmentation of planetary surface environment images.

5. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs the deep space semantic segmentation pseudo-label learning method based on error-correcting output codes as described in any one of claims 1 to 3.

6. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the deep space semantic segmentation pseudo-label learning method based on error-correcting output codes as described in any one of claims 1 to 3.

Images