CN-118212425-B - Label noise removal and spectral distribution modeling combined optimization method

Abstract

The invention provides a label noise removal and spectral distribution modeling combined optimization method, and belongs to the field of hyperspectrum. According to the method, reliable samples of the labels are extracted through non-uniform spatial sampling, constraint is carried out on classification features by utilizing a multi-center spectrum prototype, and model stable training under the condition that the labels contain noise is achieved through joint optimization of pixel-by-pixel overlapping degree loss, class cross entropy loss and spectrum prototype contrast loss. The method can dynamically remove the noise label without excessive cleaning, and the combined optimization of noise removal and spectrum modeling ensures the low noise of input data, has good model regularization effect and higher practical value.

Inventors

• YIN JIHAO
• FENG JIAQI
• JIANG HONGXIANG
• WANG QIXIONG

Assignees

• 北京航空航天大学

Dates

Publication Date

20260508

Application Date

20240423

Claims (3)

1. 1. A joint optimization method of label noise removal and spectral distribution modeling, the method comprising: Step 1, removing noise labels, namely randomly sampling from a prediction mask and a true value mask, obtaining a confidence coefficient matrix of sampling points through a confidence coefficient calculation formula, sequencing the confidence coefficient of the sampling points, and taking a plurality of points in front as label pure samples; step 2, regularizing a classification model, introducing a multi-center spectrum prototype into the classification model, carrying out similarity measurement on a vector to be classified and the multi-center spectrum prototype, determining category attribution according to a measurement result, and updating the multi-center spectrum prototype by momentum, wherein the method comprises the following steps: The record category is characterized in that Record the multi-center spectrum prototype as , wherein, For the dimension of the category characteristics, Is the first The feature vectors corresponding to the respective prediction masks, ∈ , For the number of categories to be considered, In the real number domain of the number, Is the first Multiple multi-center spectral prototype The central part of the two-way valve is provided with a plurality of centers, , Is the total number of multi-center spectrum prototypes, and each multi-center spectrum prototype has a characteristic scale of When the first Feature vectors corresponding to the prediction masks are classified into categories Feature vector and multi-center spectral prototype distances The following relationship is satisfied: , Wherein, the Is the first The feature vector corresponding to each mask is segmented A segment in which the distance measure is a vector cosine distance, expressed as The formula is as follows: , Representing the vector dot product, judging the category attribution according to the distance between the characteristic vector slice and the multi-center spectrum prototype, and when the category attribution is the first Feature vectors corresponding to the foreground prediction masks Belonging to In the case of a multi-centered spectral prototype corresponding to a class, Derived from the following formula: , Wherein, the , The method is characterized in that the method comprises the steps of respectively obtaining the class serial number and the center serial number of a multi-center spectrum prototype corresponding to the minimum feature distance measurement, wherein in the training process, the multi-center spectrum prototype adopts a momentum updating mode, and a momentum updating formula is as follows: , Wherein, the The weights are updated for the momentum and, Is the first All under class The classification feature vector center of the segment; And 3, designing an optimization loss function to remove noise labels and realize dynamic interaction iterative optimization regularized by the classification model.
2. 2. The method for joint optimization of label noise removal and spectral distribution modeling according to claim 1, wherein said step 1 comprises, The hyperspectral image of the input classification network is noted as Wherein For the number of hyperspectral bands, As the height of the image is to be taken, For the width of the image to be the same, For real number domain, hyperspectral image After encoding and decoding, the output foreground prediction mask is recorded as The truth mask is recorded as , wherein, The category number is: , , in the formula, Represent the first First of the masks The probability that a pixel point is predicted as a foreground pixel, Represent the first First of the masks The true value results for the individual pixels, ∈ Confidence matrix Element(s) Calculated from the following formula: , Representing a modulo value, wherein a dynamic confidence objective function The expression is as follows: , Wherein, the The current iteration step is indicated as such, For a set iteration step threshold value, Is a natural index.
3. 3. The method of joint optimization of label noise removal and spectral distribution modeling according to claim 2, wherein said step 3 comprises, joint optimizing a loss function Including pixel-by-pixel overlap loss Cross-class entropy loss Loss of contrast to spectral prototypes The formula is as follows: , Wherein, the , , To respectively correspond to pixel-by-pixel overlap loss Cross-class entropy loss Loss of contrast to spectral prototypes Is a loss of pixel-by-pixel overlap loss Expressed as: , class cross entropy loss Expressed as: , spectral prototype contrast loss Expressed as: , wherein N represents the total number of pixel points after screening.

Description

Label noise removal and spectral distribution modeling combined optimization method Technical Field The invention belongs to the field of hyperspectrum, and particularly relates to a label noise removal and spectral distribution modeling combined optimization method. Background The hyperspectral image has the data advantage of map unification, and has remarkable effect in the ground feature fine classification task. However, the current hyperspectral image visualization result is difficult to simultaneously display space and spectrum information, and a certain error exists in manual visual interpretation, so that the hyperspectral data tag has noise. The noise label samples can be divided into two types according to error reasons, namely an example related noise label and an example unrelated noise label, wherein the example related noise label refers to the reason and example characteristic correlation of error labeling, for example, a ground feature A and a ground feature B have similar spectrum characteristics, geographic distribution of the two have no obvious priori difference, the situation of error labeling of the ground feature A and the ground feature B exists during labeling, the example unrelated noise label refers to the reason and the example characteristic independence of error labeling, and the situation is specifically expressed as random error labeling, such as unclear ground feature contours at class boundaries, random errors existing during labeling and the like. The method has an influence on the supervision training of the hyperspectral image classification model, and seriously hinders the hyperspectral robust training process. At present, in order to solve the problem of training a hyperspectral image classification model containing label noise, the existing method can be divided into two main types, namely, label noise elimination is carried out before the model is trained, and regularization is carried out on the model to improve noise resistance. Tag noise rejection mainly refers to filtering tag noise by using priori knowledge of training samples, such as processing on extracted high-separability space-spectral features by using a limited energy minimization strategy to detect and correct noisy tags, or proposing to use a density peak clustering algorithm to identify training samples with wrong tags based on a local density decision strategy, and in this way, eliminating the training samples with wrong tags, and also constructing a spectral-space probability transfer matrix which simultaneously considers spectral similarity and spatial information based on superpixels, and eliminating interference of noise tags by using a multi-voting mode. However, the tag noise removing method has the problem of excessive cleaning, namely easy deletion of accurate real mark samples. Model regularization is a technique for preventing model overfitting, which can prevent model overfitting to tag noise during the training phase. In deep learning, due to the fact that model parameters are more, noise label data are easy to be excessively fitted in a training stage, and therefore model performance is reduced. Regularization encourages models to learn simple patterns by punishing the complexity of the model, thereby preventing the model from overfitting to noise labels and improving its generalization ability on test data. In a noise label robust training task, label smoothing reduces the supervision degree of labels on the model by reducing the probability of real labels and assigning the residual probability to other categories, and confidence regularization reduces the excessive confidence of the model on the noise labels by encouraging the model to keep cautious on the output results during training, for example, by introducing additional loss terms through KL divergence, and constrains the output features. However, model regularization is more effective when the noise label proportion is lower, and if the noise label quantity is too large, the noise-resistant effect of model regularization is greatly reduced. Disclosure of Invention In order to solve the technical problems, the invention provides a label noise removal and spectrum distribution modeling combined optimization method, which extracts a label reliable sample through non-uniform spatial sampling, utilizes a multi-center spectrum prototype to constrain classification characteristics, and realizes model stable training under a label noise condition through combined optimization of pixel-by-pixel overlapping degree loss, category cross entropy loss and spectrum prototype contrast loss. The method can dynamically remove the noise label without excessive cleaning, and the combined optimization of noise removal and spectrum modeling ensures the low noise of input data, has good model regularization effect and higher practical value. In order to achieve the above purpose, the present invention proposes the following technical scheme: A joint o