CN-121982406-A - Multi-label remote sensing image classification method based on pseudo-random Top-K ordering

CN121982406ACN 121982406 ACN121982406 ACN 121982406ACN-121982406-A

Abstract

The invention discloses a multi-label remote sensing image classification method based on pseudo-random Top-K sorting, which comprises the steps of constructing a multi-label sorting classification model, processing a multi-label sorting task of a remote sensing image through the multi-label sorting classification model, executing pseudo-random sorting through the multi-label sorting classification model, calculating soft Top-K normalized break accumulation gain, performing transitional Top-K optimization training, integrating the pseudo-random sorting, the soft Top-K normalized break accumulation gain and the transitional Top-K optimization through the multi-label sorting classification model, constructing a pseudo-random Top-K sorting classification frame, introducing a neutral label into the pseudo-random Top-K sorting classification frame, restricting a neutral label predicted value to an interval between a positive label predicted value and a negative label predicted value, optimizing the pseudo-random Top-K sorting loss, training the multi-label sorting classification model through a data set, inputting the trained multi-label sorting classification model into the remote sensing image, and outputting a remote sensing image sorting result through the multi-label sorting classification model.

Inventors

Zhong Ruiru
LIU YISHU
HUANG HAORAN
ZHENG XIAODONG
ZHANG ZHANPENG
Zhou Diewei

Assignees

华南师范大学

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (10)

1. A multi-label remote sensing image classification method based on pseudo-random Top-K ordering is characterized by comprising the following steps: constructing a multi-label sorting classification model, and processing multi-label sorting tasks of the remote sensing image through the multi-label sorting classification model; Performing pseudo-random sorting through a multi-label sorting classification model, calculating soft Top-K normalization break accumulated gain, and performing transitional Top-K optimization training; The multi-label sorting classification model integrates pseudo-random sorting, soft Top-K normalization break accumulation gain and transitional Top-K optimization, and a pseudo-random Top-K sorting classification framework is constructed; The pseudo-random Top-K sorting framework introduces neutral labels, the predicted values of the neutral labels are constrained in the interval between the predicted values of the positive labels and the predicted values of the negative labels, and the pseudo-random Top-K sorting loss is optimized; training the multi-label sorting classification model by using the data set, inputting the trained multi-label sorting classification model into the remote sensing image, and outputting a remote sensing image sorting result by the multi-label sorting classification model.
2. The method of classifying multi-tag remote sensing images based on pseudo-random Top-K ordering of claim 1, wherein performing pseudo-random ordering comprises: when the image is input into the model, the model outputs a label predictive score vector ; Defining a loss function The loss function is a function of the ranking index Space of possible ordering Desired value of (a) wherein And weighted according to their probability under PL distribution, the formula is as follows: ; Wherein, the Representation ordering A kind of electronic device Distribution probability; Representation of True relevance scores; is a tiny substitution loss function used for measuring the predictive ranking And by The gap between the determined ideal ordering; Representing all Corresponding to Sum of products of (2); In the context of large-scale aerial images, the number of labels and images is enormous, so that a random simulation method is adopted, namely random extraction from uniform distribution is adopted The independent samples are converted into Gumbel noise disturbance, and the log fraction after disturbance is constructed to obtain The ordered sequences and averaging the resulting substitution losses; Is provided with Is the first The log fraction of the sub-random disturbance after Gumbel noise disturbance can obtain approximate loss: ; Wherein, the Representing the approximate loss value obtained using a random simulation method; Representing the ordered sequence obtained after Gumbel noise disturbance.
3. The method of classifying multi-tag remote sensing images based on pseudo-random Top-K ordering of claim 2, wherein performing pseudo-random ordering comprises: To address the defect of introducing random sampling, the front in deterministic low-difference sequence is adopted Instead of randomly decimated points A sample number; Is provided with Is the first The final approximate loss is then the following deterministic log score based on low-variance sequence perturbation: ; Wherein, the Representing the approximate loss value obtained using a pseudo-random ordering method; Representing the ordered sequence obtained after perturbation with a low differential sequence.
4. The method for classifying a multi-label remote sensing image based on pseudo-random Top-K ordering of claim 3, wherein calculating soft Top-K normalized break cumulative gain comprises: The normalized broken cumulative gain is a standard ordering index in information retrieval and is used for measuring the coincidence degree of the ordering generated by the model and the ideal correlation ordering Representing a set of all possible orderings, Representation model generation The number of the generalized scores is the number, Representation of The true relevance score is a score of the true relevance, Representation press The index order in the descending order of arrangement, The representation is composed of The determined ideal ordering has a compromise cumulative gain: ; Wherein, the Representing a break cumulative gain between the predicted ordering and the ideal ordering; represent the first Items; Representation of Is a correlation score of (2); Normalizing the damage accumulated gain to obtain normalized damage accumulated gain: ; Wherein, the Representing normalized break cumulative gain between predicted and ideal ordering when When NDCG@N is a list index evaluating complete ordering, when When NDCG@K is used for evaluating the first K orders, the larger the value is in the range of (0, 1) of the value range of NDCG@K, the higher the ordering quality is.
5. The method for classifying multi-label remote sensing images based on pseudo-random Top-K ordering according to claim 4, wherein calculating soft Top-K normalized break cumulative gain comprises: The StopNDCG@K is provided, and the method of directly approaching the NDCG@K in a differential mode is adopted based on generalized scores Instead of "hard" allocation of the correlation gain at discrete locations in the ordered list; Using soft probability distribution Substitute for independent heat vector Estimating the probability of which item belongs to the kth position, further obtaining the approximate value of the relevant gain at the kth position, then updating the score vector for restraining the influence of the item which has been 'soft-selected', and obtaining the slightly approximate value of DCG@K by repeatedly executing the soft allocation and restraining process for K times, which is marked as StopDCG@K: ; ; Wherein, the An approximation representing the correlation score at the kth position resulting from the soft allocation; Representing a slightly approximated dcg@k between the predicted ordering and the ideal ordering after the soft allocation process; Normalizing StopDCG@K to obtain a final normalized measurement StopNDCG@K: ; Wherein, the The value range of StopNDCG@K is also (0, 1), and the larger the value is, the higher the ordering quality is; Thus, the loss function is defined as follows: ; Wherein, the The soft Top-K normalized break cumulative gain is directly optimized by treating the entire ordering as a single structured output.
6. The pseudo-random Top-K ordering based multi-tag remote sensing image classification method according to claim 5, wherein performing transitional Top-K optimization training comprises: Transitional Top-K is optimized as a learning strategy that can dynamically fuse two different cut-off targets by first learning an easier subtask, namely shallow cut-off value Top-K 1 , then gradually transitioning to a more challenging final target, namely deep cut-off value Top-K 2 , where K 1 <K 2 ; transient Top-K optimized loss function The convex combination defined as two soft Top-K normalized break cumulative gain losses: ; Wherein, the Representing the loss of the shallow cut-off value top-K 1 ; Representing the loss of the deep cut-off value top-K 2 , the mixing coefficient The scheduling function of the training iteration t is designed to monotonically decay from 1 to 0, and as training progresses, the optimal focus is seamlessly switched from Top-K 1 to the Top-K 2 target with higher requirements.
7. The multi-label remote sensing image classification method based on pseudo-random Top-K ordering according to claim 6, wherein the pseudo-random Top-K ordering classification framework introduces a neutral label, and the neutral label predicted value is constrained in an interval between a positive label predicted value and a negative label predicted value, specifically: Given tag space Order-making A positive set of labels representing sample x, Representing a negative set of labels, the task of the model is to learn a mapping from sample x to n label ranks, when sample x is input into the model, n scores s are output 1 ,s 2 ,……,s n Wherein s i represents Probability of positive label correspondence Negative label correspondence ; To enhance the distinction between positive and negative labels, a method of A plurality of neutral labels, wherein the preset fraction of the neutral labels is fixed in the interval Wherein Thereby forming a buffer band so that the predictive score of the positive label is concentrated The predictive score of negative labels is focused on 。
8. The multi-label remote sensing image classification method based on pseudo-random Top-K ordering according to claim 7, wherein the optimizing of pseudo-random Top-K ordering loss is specifically: constructing a score vector by appending a fixed "neutral score" to the model predictive score : ; Wherein S i represents the i-th prediction score in the score vector output by the model; corresponding correlation vector The definition is as follows: ; Subsequently, in a pseudo-random ordering paradigm, generic micro-substitution loss is possible The soft Top-K normalized break cumulative gain is used to normalize the loss function deduced by the break cumulative gain and transitional Top-K optimization, i.e. the soft Top-K normalized break cumulative gain loss is fused ; Thus, the resulting pseudo-random Top-K ordering penalty Expressed as: ; Wherein, the By pairing in a pseudo-random ordering process And performing the sorting generated by the j-th deterministic low-difference sequence disturbance.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the pseudo-random Top-K ordering based multi-label remote sensing image classification method of any one of claims 1 to 8.
10. A storage medium storing a computer program, wherein the computer program when executed by a processor implements the pseudo-random Top-K ordering based multi-label remote sensing image classification method of any one of claims 1 to 8.

Description

Multi-label remote sensing image classification method based on pseudo-random Top-K ordering Technical Field The invention relates to the technical field of remote sensing image classification, in particular to a multi-label remote sensing image classification method based on a list ordering strategy and pseudo-random Top-K ordering. Background With the rapid development of image acquisition technologies such as unmanned aerial vehicles and high-resolution satellites, ground object targets in aerial images show remarkable characteristics of various scales, complex distribution and overlapped semantics, and the traditional single-tag classification method cannot meet the actual requirements of one-figure multi-class, so that the multi-tag classification technology becomes a research hot spot. Early multi-label aerial image Classification (ML-AIC) methods typically combined hand-made features with traditional machine learning models. However, in the last decade, deep neural networks have become the dominant paradigm. The current mainstream multi-label aviation image classification methods are mostly constructed based on Binary Cross Entropy (BCE) loss functions, and the core focuses on Binary judgment of label presence or absence, but ignores the obvious difference between visual weight and semantic importance occupied by different ground objects in images. The lack of structural information directly leads to weak model generalization capability, and classification accuracy is greatly reduced in complex scenes. Technical researches in the field of multi-label aviation image classification mainly surround two major forms of classification optimization and sequencing optimization. Multi-label classification techniques for trend classification optimization. The multi-label classification technology based on the binary cross entropy is a main stream technology in the field of current multi-label aviation image classification, and the core thought is to disassemble a multi-label classification task into a plurality of independent two-class tasks, and the judgment capability of 'presence/absence' of each label is realized through a binary cross entropy loss function optimization model. Essentially, these mainstream methods focus on tag recognition, with the optimization objective still correctly recognizing positive tags, and ignoring the semantic ordering structure inherent in visual scenarios, which limits its versatility and robustness. Multi-label classification techniques for trend ranking optimization. In order to solve the problem that the binary cross entropy method ignores label semantic ordering, researchers introduce a multi-label ordering (Multilabel Ranking, MLR) paradigm, and semantic priority relation optimization among focused labels is mainly subdivided into three types of technologies of 'point-by-point ordering', 'paired ordering' and 'list ordering'. The point-by-point ranking technique assigns scores to individual tags by learning the utility function of each tag, and then ranks the tags according to the scores. Such methods typically ignore inter-tag dependencies, thereby affecting overall effectiveness. The core of the pairwise ordering technique is to optimize the relative order of tag pairs, creating an overall ranking by determining the relative relevance between each pair of tags. However, its optimized loss function often does not agree with the standard ranking indicators, resulting in disjoint learning objectives and ranking quality. Furthermore, training is often computationally expensive due to the quadratic increase in the number of tag pairs—this is especially true for multi-tag aerial image classification tasks where the number of tags is typically large. The list ordering technology takes the ordering quality of the complete tag sequence as an optimization target, and improves the overall semantic consistency by optimizing the whole sequence ordering index (such as normalized break cumulative gain). The method can directly alleviate the problem of target mismatch in the paired method, so the method is generally superior to the point-by-point ordering and paired ordering methods in terms of prediction performance and calculation efficiency. More fundamentally, it is assumed that by considering the entire tag set as a structured output, the list approach is able to essentially discern subtle correlation differences between tags, thereby effectively mining fine-grained ordering information that cannot be provided by binary positive and negative supervision alone. Despite the broad prospects of multi-tag ranking based methods, this field still faces significant challenges. First, the tabular ordering technique in multi-tag ordering has largely not been fully explored, meaning that the potential of the tabular approach to capture fine-grained dependencies between tags has been hardly examined in multi-tag ordering. Second, all existing multi-label ranking models are deterministic in n