Semi-supervised change detection framework combining double-branch self-knowledge distillation and multi-granularity disturbance enhancement

Abstract

The invention discloses a semi-supervised change detection framework combining double-branch self-knowledge distillation and multi-granularity disturbance enhancement. Comprising the following steps: designing a semi-supervised learning training framework of double-branch self-knowledge distillation; the output of the undisturbed branch is used as a teacher model, and the output set generated by the multi-granularity disturbance enhancement branch forms a student model set; teacher and student branch weight sharing is realized, and bidirectional knowledge distillation is realized by means of an independent decoder; a knowledge distillation design loss function is introduced for model training. The method is almost comparable with full supervision performance, a beneficial insight is provided for a semi-supervised learning normal form of geographic space artificial intelligence, and experimental results on a plurality of public data sets show that compared with the previous best result, the performance is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of semi-supervised remote sensing image change detection, and particularly relates to a semi-supervised change detection framework combining double-branch self-knowledge distillation and multi-granularity perturbation enhancement. BACKGROUND

[0002] Change detection is a core task in remote sensing image analysis, aiming to identify surface changes by comparing multi-temporal images of the same area, and has important applications in urban planning, disaster assessment and environmental monitoring. Although supervised deep learning-based change detection methods have shown excellent performance in the interpretation of dual-temporal high-resolution remote sensing images, their effectiveness is highly dependent on a large amount of pixel-level labeled data. However, manually labeling dual-temporal images through pixel-level change comparison is extremely expensive and time-consuming, which severely limits its practical application. This challenge has driven the research of semi-supervised change detection methods, which utilize limited labeled data and abundant unlabeled samples.

[0003] Existing semi-supervised change detection methods mainly focus on consistency regularization and pseudo-label optimization. For example, C2F-SemiCD proposes a coarse-to-fine consistency training framework, which uses an average teacher model to generate pseudo-labels for unlabeled data. SemiCD enhances feature consistency constraints by perturbing feature difference maps, while MTCNet innovatively introduces single-temporal semantic supervision and building priors to improve training efficiency. RCCD adopts a dual-stream network and spatial-feature hybrid perturbation to enhance the robustness of the model. Although progress has been made, current methods still face two persistent bottlenecks: (1) Noise sensitivity: Traditional image-level perturbations (e.g., rotation, cropping) are easily affected by irrelevant noise such as light changes and cloud cover, reducing the reliability of pseudo-labels; (2) Insufficient discriminability: Single-granularity perturbation strategies struggle to cope with multi-dimensional change patterns such as spatial blurring and feature deactivation, limiting their adaptability to complex scenarios.

[0004] To address these challenges, the present application proposes a semi-supervised change detection framework based on double-branch self-knowledge distillation and multi-granularity perturbation enhancement. The present application introduces a teacher-student network co-evolution mechanism: the teacher network generates temperature-calibrated soft labels to preserve the probability distribution information of unlabeled data; the student network applies double-granularity perturbation, including random blurring in image space (simulating imaging noise) and channel dropout in feature space (simulating feature degradation). SUMMARY

[0005] The application provides a semi-supervised change detection framework combined with double-branch self-knowledge distillation and multi-granularity perturbation enhancement.

[0006] Step 1: a double-branch self-knowledge distillation semi-supervised learning training framework is designed;

[0007] Step 2: the output of the non-perturbation branch is used as a teacher model, and the output set generated by the multi-granularity perturbation enhancement branch constitutes a student model set;

[0008] Step 3: the teacher and student branch weights are shared, and at the same time, bidirectional knowledge distillation is realized by relying on an independent decoder;

[0009] Step 4: a knowledge distillation loss function is introduced for model training.

[0010] The semi-supervised change detection framework combined with double-branch self-knowledge distillation and multi-granularity perturbation enhancement has the following advantages compared with the traditional semi-supervised change detection method:

[0011] (1) the application uses Gaussian kernel convolution to simulate imaging degradation at the image level, and adopts a channel dropout strategy to randomly shield non-key feature channels at the feature level.

[0012] (2) the feature perturbation operation is applied in the 3D tensor space at the same time, and compared with the traditional spatial dropout which mainly operates in the spatial dimension, the channel dimension constraint is introduced, and by adjusting the value of p*, different intensity of perturbation feature tensors can be generated. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The semi-supervised change detection framework combined with double-branch self-knowledge distillation and multi-granularity perturbation enhancement is shown in the schematic diagram. DETAILED DESCRIPTION

[0014] The method described in the application is described in detail in combination with the drawings and examples.

[0015] The semi-supervised change detection framework combined with double-branch self-knowledge distillation and multi-granularity perturbation enhancement is shown in the schematic diagram. Figure 1The method is shown. The overall process is as follows: first, generate a student model set through multi-granularity perturbation enhancement, input the original data set into the teacher model, generate pseudo annotation for the unmarked data by the teacher model, share the weights of the teacher and student branches, and realize bidirectional knowledge distillation by relying on independent decoders. The knowledge distillation stage defines three types of loss functions to optimize parameters, and the output of the teacher network guides the learning of the student network.

[0016] Step 1: design a semi-supervised learning training framework of double-branch self-knowledge distillation;

[0017] In designing the framework, we first construct two network branches that share the parameters of the encoder but have independent decoders, namely the teacher network and the student network. Such a structure design not only ensures the commonality of the two in feature extraction, but also gives the difference in the decoding process, providing a foundation for subsequent knowledge distillation.

[0018] Step 2: the output of the branch without perturbation serves as the teacher model, and the output set generated by the multi-granularity perturbation enhancement branch constitutes the student model set; generate a student model set through multi-granularity perturbation enhancement, input the original data set into the teacher model, and generate pseudo annotation for the unmarked data by the teacher model. The weak perturbation at the image level and the feature level serves as the assistant model for the corresponding strong perturbation.

[0019] Step 3: share the weights of the teacher and student branches, and realize bidirectional knowledge distillation by relying on independent decoders;

[0020] The loss between the teacher model and the student model set is measured as formula (1), and the loss between the assistant model and the student model set is measured as formula (2),

[0021]

[0022] Step 4: introduce a knowledge distillation designed loss function for model training.

[0023] The semi-supervised training stage integrates the pre-training representation of the supervision stage in layers, and optimizes the model on labeled and unlabeled data through consistency regularization terms. Specifically, the loss function combines task-specific supervision and perturbation-driven consistency alignment, and the calculation is as formula (3)

[0024]

Claims

1. A semi-supervised change detection framework combining dual-branch self-knowledge distillation and multi-granularity perturbation enhancement, characterized by: Through a dual-branch self-knowledge distillation framework, the calibrated soft labels generated by the teacher network are used to guide the student network's predictions, effectively reducing the accumulation of pseudo-label noise and improving generalization capabilities. At the same time, a multi-granularity perturbation enhancement strategy coordinates spatial and feature perturbations to enhance the model's adaptability and robustness to complex scenarios, thereby achieving high performance in semi-supervised remote sensing change detection. Specifically, the following steps are involved: 1) Design a two-branch semi-supervised learning training framework for self-knowledge distillation; 2) The output of the unperturbed branch serves as the teacher model, and the output set generated by the multi-granularity perturbation enhancement branch constitutes the student model set; 3) The teacher and student branches share weights, while relying on independent decoders to achieve bidirectional knowledge distillation; 4) Introduce knowledge distillation to design loss function for model training.

2. The semi-supervised change detection framework combining dual-branch self-knowledge distillation and multi-granularity perturbation enhancement as claimed in claim 1, characterized in that The teacher network dynamically generates adaptive soft targets by fusing the deep features and predicted distribution of the student network in real time, replacing the parameter caching mechanism of the traditional momentum teacher network.

3. The semi-supervised change detection framework combining dual-branch self-knowledge distillation and multi-granularity perturbation enhancement as claimed in claim 1, characterized in that At the image level, Gaussian kernel convolution is used to simulate imaging degradation (such as cloud occlusion and sensor noise); at the feature level, channel dropout strategy is used to randomly mask non-critical feature channels.

4. The semi-supervised change detection framework combining dual-branch self-knowledge distillation and multi-granularity perturbation enhancement as claimed in claim 1, characterized in that The feature perturbation operation applies a mask simultaneously in the 3D tensor space. Compared with the traditional spatial dropout that mainly operates on the spatial dimension, it introduces the constraint of the channel dimension. By adjusting the value of p*, perturbation feature tensors of different intensities can be generated.

5. The semi-supervised change detection framework combining dual-branch self-knowledge distillation and multi-granularity perturbation enhancement as claimed in claim 1, characterized in that The loss function combines task-specific supervision and perturbation-driven consistency alignment:

Images