Generative Artificial Intelligence-Driven Image Semantic Segmentation and Visual Scene Understanding for English Language Teaching Environments

2.1 Overall framework overview

GSP-Net adopts pre-trained Stable Diffusion 2.1 as the backbone to construct an end-to-end image semantic segmentation and visual scene understanding framework oriented toward English teaching scenarios. Let the input teaching scene image be $I \in \mathrm{R}^{H \times W \times 3}$; the network outputs a pixel-level semantic segmentation mask M, a structured scene graph $G_{scene}$, and a teaching-semantic-enhanced image $I_{gen}$. The overall framework consists of three collaborative modules: a teaching-context-driven dynamic prompt encoding and hierarchical attention anchoring module, which injects scene semantic priors into the diffusion process; a spatial-density-driven multi-scale attention topology fusion module, responsible for generating refined segmentation masks; and a differentiable scene graph propagation and generative closed-loop feedback module, which completes the cross-modal mapping from segmentation to semantic reasoning and enables self-supervised evolution. The three modules share a computational graph during back-propagation, forming a unified system that can be optimized end-to-end.

The U-Net denoising sub-network of the diffusion model produces two complementary types of attention maps during generation. Let the number of U-Net decoder layers be L, at denoising time step t, the cross-attention map of the l-th layer is $A_{l}^{cross, t} \in \mathrm{R}^{H_l \times W_l \times L_p}$, where $H_l \times W_l$ is the spatial resolution and $L_p$ is the text prompt length. The self-attention map is A^self,t_l $\in \mathrm{R}^{H_l \times W_l \times D_l}$, where $H_l \times W_l$, where $D_l$ is the number of attention heads at that layer. The cross-attention maps characterize the response intensity between each semantic category in the text condition and the image spatial positions, while the self-attention maps encode the long-range dependencies among pixels within the image and the shape-topological information of objects. The core design of GSP-Net lies in the scene-adaptive extraction, fusion, and enhancement of these two types of attention maps, rather than adopting the simple averaging or fixed-weight combination strategies used in existing methods. A differentiable attention gating network dynamically adjusts the contribution ratio of self-attention at each layer based on spatial location and local image characteristics, allowing the fusion process to adaptively adjust with scene content. A text density enhancement factor further focuses on text-dense regions such as writing areas and projection zones, locating high-texture directional regions via structure tensor analysis and amplifying their attention responses, effectively alleviating semantic adhesion at dense text boundaries.

The inference stage adopts a single-pass denoising strategy to balance computational efficiency and segmentation quality. Given the input image I and the scene-aware condition $C_{\hat{t}}$, the model executes only T = 20 denoising steps instead of the conventional 50-step full sampling of diffusion models. This design is based on the following observation: attention maps have already converged to a stable semantic response distribution in the early-to-mid denoising steps; further increasing the number of steps yields only marginal improvement in segmentation accuracy while significantly increasing computational burden. To eliminate random fluctuations in single-step attention maps, attention maps across all time steps are aggregated to obtain stable semantic responses:

$\bar{A}_l^{cross}=\frac{1}{T} \sum_{t=1}^T A_l^{cross, t}, \bar{A}_l^{self}=\frac{1}{T} \sum_{t=1}^T A_l^{self,t}$,                (1)

where $\bar{A}_l^{cross}$ and $\bar{A}_l^{self}$ are the time-step-aggregated cross-attention maps and self-attention maps, respectively, serving as inputs for subsequent fusion and enhancement processing. The total computational complexity of GSP-Net is $O(T \cdot L \cdot H \cdot W \cdot d)$, where $d$ is the feature dimension of attention heads. Compared with baseline methods, the additionally introduced computational overhead mainly comes from the forward propagation of the differentiable gating network $(O(L \cdot H \cdot W))$ and the computation of the structure-tensor-based text density factor ($O\left(H \cdot W \cdot k^2)\right.$), where $k$ is the Gaussian kernel size); the combined portion accounts for approximately 12% to 15% of total inference time, which is within an acceptable range for practical deployment.

2.2 Teaching-context-driven dynamic prompt encoding and hierarchical attention anchoring

Semantic objects in teaching scenarios possess a distinct hierarchical structure, progressing from physical facilities to instructional media and further to abstract semantic content, with a clear concrete-to-abstract progressive relationship. The general category systems adopted by existing zero-shot segmentation methods struggle to characterize this hierarchical semantic organization, resulting in non-targeted responses in cross-attention maps over semantic categories. To address this, we construct the English Teaching Scene Hierarchical Semantic Category System, named it as the English Teaching Scene Tree (ETST), dividing all semantic concepts into three tiers: the physical environment layer covers eight categories of basic physical facilities including blackboard, podium, desks, projection screen, wall, and window; the instructional media layer contains seven categories of carriers bearing teaching information, such as writing text areas, projection content areas, word cards, textbook pages, exercise books, and electronic screens; the semantic content layer includes six categories of semantic units directly related to core teaching content, including English vocabulary, sentence structures, dialogue bubbles, grammatical annotations, and phonetic symbol annotations. Let the complete category set be $C=C_{L 1} \cup C_{L 2} \cup C_{L 3}$, with a total of 21 categories. The semantic focus involved in different lesson types varies significantly: vocabulary lessons emphasize vocabulary and phonetic symbol regions, while reading lessons additionally focus on sentence structures and dialogue bubbles. Therefore, the activation status of each category in the model changes dynamically with the scene type rather than being fully enabled. Figure 1 illustrates the teaching-context-driven dynamic prompt encoding and hierarchical attention anchoring mechanism.

1.png

Figure 1. Teaching-context-driven dynamic prompt encoding and hierarchical attention anchoring mechanism

To achieve scene-adaptive semantic anchoring, lesson-type discrimination is first performed on the input image. A lightweight EfficientNet-B0 is adopted as the scene classifier $f_{scene}$ to determine the lesson-type category to which the current teaching scene belongs:

$\begin{gathered}\hat{t}=\arg \max _{t_k \in T} P\left(t_k \mid I\right), \\ T=\{\text {vocabulary, reading, listening & speaking, grammar}\},\end{gathered}$               (2)

where $P\left(t_k \mid I\right)$ is the class probability distribution output by the classifier. This classifier achieves a classification accuracy of 98.7% on the self-built dataset, providing a reliable basis for subsequent dynamic semantic screening. Based on the classification result $\hat{t}$, a semantically relevant subset $C_{\hat{t}}^{select} \subseteq C$ is selected from the complete category set, so that the crossattention of the diffusion model only needs to perform matching within the semantic space corresponding to relevant categories, effectively suppressing response interference from irrelevant semantic regions. Unlike directly using manually set fixed text prompts, this method maps the selected semantic categories into a continuous space via a lightweight adapter to obtain conditional embeddings:

$C_{\hat{t}}=Adapter\left(\right.Embed\left.\left.(C_{\hat{t}}^{select}\right.)\right.) \in \mathrm{R}^d$,               (3)

where, Embed is a fixed word embedding layer, Adapter consists of a two-layer multilayer perceptron (MLP) with dimension mapping from d to 4d and back to d, the value of d takes 768, and the introduced parameter count is approximately 2.4 million. This continuous prompt embedding replaces traditional text descriptions as the cross-modal conditional input to the U-Net of the diffusion model, preserving original category semantic information while gaining tunability in the continuous space.

During the decoding process of the U-Net, the computation method of cross-attention maps at each layer directly affects the spatial localization accuracy of semantic categories. The cross-attention map of the l-th layer decoder is given by:

$A_l^{cross}(i, j)=Softmax\left(\frac{\mathrm{Q}_l(i, j) \cdot K_l^T\left(C_{\hat{t}}\right)}{\sqrt{d_k}}\right)$,                (4)

where, $Q_l \in \mathrm{R}^{H_l W_l \times d_k}$ is the query matrix obtained by linearly projecting image features, $K_l \in R^{\mid celect_t \mid \times d_k}$ is the key matrix obtained by key mapping of the conditional embedding $C_{\hat{t}}$, and $d_k=64$ is the dimension of the key space. The underlying logic of this formulation is to match semantic hierarchies with spatial receptive fields: categories at the L3 semantic content layer, such as English vocabulary and sentence structures, cooperate with self-attention in deeper decoders, whose larger receptive fields can capture semantic associations among word regions scattered across the image; categories at the L1 physical environment layer, such as blackboard edges and podium contours, combine with self-attention in shallower decoders, whose smaller receptive fields benefit the accurate localization of geometric boundaries of physical facilities. This hierarchical anchoring design allows cross-attention maps to specialize at different scales, avoiding mutual interference between high-level semantics and low-level details under a single scale.

In teaching scenes, certain categories tend to co-occur; for example, blackboards and writing text areas exhibit high co-occurrence, as do projection screens and projection content areas. Let the co-occurrence frequency of category p and category q in the training set be $f_{p q}$; then the calibrated cross-attention response is:

$\begin{gathered}A_l^{{cross,cal }}(i, j, p)= A_l^{ {cross }}(i, j, p) \cdot\left(1+\lambda_{ {co }} \sum_{q \neq p} f_{p q} \cdot I\left(\hat{y}_q(i, j)>\tau_{ {conf }}\right)\right),\end{gathered}$                (5)

where, $\lambda_{c o}=0.3$ is the co-occurrence enhancement coefficient, $\tau_{{conf }}=0.5$ is the confidence threshold, $I(\cdot)$ is the indicator function, and $\hat{y}_q(i, j)$ is the current predicted confidence of category q at pixel (i,j). This calibration takes effect at the pixel level: when a pixel shows a medium-confidence response to category p and a high-confidence response of a highly co-occurring category q exists in its neighborhood, the response of category p is enhanced; conversely, if a region isolatesly responds to a category rarely co-occurring with others in teaching scenes, its confidence is suppressed. This mechanism utilizes the inherent structure of scene semantics as a regularizer, effectively reducing isolated false segmentation caused by local texture ambiguity.

2.3 Spatial-density-driven multi-scale attention topology fusion

Self-attention maps output by different layers of the diffusion U-Net decoder exhibit essential differences in semantic granularity: self-attention maps of shallow decoders have higher spatial resolution and are sensitive to local edges and texture details; self-attention maps of deep decoders undergo multiple downsampling steps, with significantly enlarged receptive fields, capturing overall object shapes and global spatial layouts. Existing methods apply simple averaging or fixed weighted combinations to self-attention maps at all layers, overlooking the spatial-position-dependent variation in the need for detailed versus global information. In English teaching scenarios, character-level boundaries in writing areas require precise localization by shallow high-resolution features, while the distinction between projection content areas and their surrounding environment relies more on semantic context provided by deep features. To this end, self-attention maps A^self,t_l $\in \mathrm{R}^{H_l \times W_l \times D_l}$ are extracted from all L=4 decoder layers of the U-Net. The spatial resolutions of each layer are 64 × 64, 32 × 32, 16 × 16, and 8 × 8, with corresponding attention head counts $D_l$ being 8, 8, 16, and 16 in layer order. All self-attention maps are upsampled to a unified resolution H × W = 64 × 64 via bilinear interpolation:

$\widehat{A}_l=BilinearUpsample \left(A_l^{self}\right) \in \mathrm{R}^{H \times W \times D_l}$,                (6)

So that attention responses at different scales can be compared and fused pixel-wise on the same spatial grid. Figure 2 illustrates the architecture of the spatial-density-driven multi-scale attention topology fusion network, a Differentiable Attention Gating (DAG).

2.png

Figure 2. Spatial-density-driven multi-scale attention topology fusion network (Differentiable Attention Gating, DAG)

After obtaining multi-level self-attention maps at a unified spatial resolution, the key lies in determining the contribution weight of each layer at every spatial position. Unlike global fixed weights or simple averaging strategies adopted by existing methods, a DAG network is designed to achieve spatial-location-aware dynamic fusion. The network computes fusion weights $\alpha_l(i, j)$ for each decoder layer at each spatial position, whose values are jointly determined by three factors: the intensity of the self-attention response at that position, whether the position belongs to a text-dense region, and the structural similarity of attention maps between adjacent layers. The weights adopt a Softmax-normalized form:

$\alpha_l(i, j)=\frac{\exp \binom{\phi_l \cdot \bar{A}_l(i, j)+\beta_l \cdot I(D(i, j)>\tau)}{+\gamma_l \cdot \operatorname{Sim}\left(\bar{A}_l, \bar{A}_{l-1}\right)}}{\sum_{m=1}^L \exp \binom{\phi_m \cdot \bar{A}_m(i, j)+\beta_m \cdot I(D(i, j)>\tau)}{+\gamma_m \cdot \operatorname{Sim}\left(\bar{A}_m, \bar{A}_{m-1}\right)}}$,             (7)

where, $\bar{A}_l(i, j)=1 / D_t \sum_{d=1}^{D_l} \widehat{A}_l(i, j, d)$ is the mean response of the $l$-th layer self-attention map averaged over the attention-head dimension, reflecting the overall activation intensity of that layer at the position. $D(i, j)$ is the text density factor, $\tau$ is the density threshold, and the indicator function $I(D(i, j)>\tau)$ marks whether the pixel lies in a text-dense region. $\operatorname{Sim}\left.(\bar{A}_l\right., \bar{A}_{l-1})$ is the cosine similarity between attention maps of adjacent layers, measuring consistency in their spatial response distributions. $\phi_l, \beta_l, \gamma_l$ are learnable scalar parameters for each layer, controlling the contributions of base response intensity, text-region enhancement magnitude, and inter-layer consistency constraint, respectively. The entire gating network introduces only 3L = 12 learnable parameters, achieving adaptive adjustment of fusion weights with almost no increase in model capacity.

The dense text arrangement in writing and projection areas of English teaching scenes imposes special requirements on segmentation boundary precision. Edge detection operators commonly used in general scene segmentation methods struggle to effectively distinguish densely arranged character edges from true object boundaries, leading to character adhesion in writing areas and blurred boundaries between projected content and background. A structure-tensor-based text density metric is introduced, encoding texture structure information within each pixel neighborhood into a second-order matrix. For pixel (i,j), its structure tensor is defined as:

$J_\rho(i, j)=G_\rho *\left(\nabla I_\sigma(i, j) \nabla I_\sigma(i, j)^{\top}\right)$,                  (8)

where, $I_\sigma$ is the Gaussian-smoothed image with smoothing parameter $\sigma=0.8, G_\rho$ is a Gaussian kernel with scale parameter $\rho=1.5$, and $\nabla=\left(\partial_x, \partial_y\right)^T$ is the spatial gradient operator. The structure tensor $J_\rho$ is a positive semi-definite second-order symmetric matrix, whose eigenvalues $\lambda_1(i, j)$ and $\lambda_2(i, j)$ reflect the anisotropic distribution of gradient energy in the pixel neighborhood. The text density factor is defined as the determinant of the structure tensor:

$D(i, j)=\operatorname{det}\left(J_\rho\right)=\lambda_1(i, j) \cdot \lambda_2(i, j)$               (9)

In text-dense regions, character strokes produce significant gradient responses along multiple directions, keeping both eigenvalues at relatively high levels and raising the determinant response markedly; in smooth regions such as walls or podium surfaces, gradient responses are weak in all directions, and the determinant approaches zero. This metric is rotation-invariant to texture orientation, thus stably responding to English writing oriented horizontally, vertically, or obliquely. In statistics from the self-built dataset, the mean D in writing areas is approximately 8.3 times that of smooth regions, indicating good discriminability of this factor.

Based on the above design, multi-level self-attention maps are fused according to the weights output by the DAG network, and the fusion result is spatially adaptively enhanced by introducing the text density factor. The fusion process takes the form:

$M_{ {fused }}(i, j)=\frac{\sum_{l=1}^L \alpha_l(i, j) \cdot \bar{A}_l(i, j)-\min _{(u, v)}(A)}{\max _{(u, v)}(A)-\min _{(u, v)}(A)} \times\left(1+\gamma \cdot \frac{D(i, j)}{\max _{(u, v)}(D)}\right)$                (10)

The first term is the weighted fused self-attention response min-max normalized to the [0, 1] interval, and the second term is the normalized enhancement factor derived from text density, with γ = 0.3 as the enhancement coefficient. The mechanism of this formulation is as follows: in text-dense regions such as writing and projection areas, $D(i, j)$ takes larger values, and the fused response is locally amplified, yielding sharper localization of segmentation boundaries; in smooth regions such as walls and podiums, $D(i, j)$ approaches zero, the enhancement factor tends to one, and the fused response is virtually unaffected, preventing spurious responses introduced by over-enhancement in textureless areas. The enhanced self-attention map is then taken as a spatial structural prior and multiplied element-wise with the calibrated cross-attention maps to obtain the final confidence map for each semantic category:

$M_p^{ {final }}=M_{ {fused }} \odot A^{ {cross }, { cal }}(p), \forall p \in C_{\hat{t}}^{ {select }}$                  (11)

where, ⊙ denotes the Hadamard product. The cross-attention map provides the semantic probability response of a pixel belonging to category p, while the self-attention map provides structural integrity constraints at that location; their product ensures segmentation results are semantically accurate while possessing spatial topological plausibility. The final semantic category of each pixel is determined by taking the maximum response among all activated categories:

$\hat{y}(i, j)=\arg \max _{p \in C_{\hat{t}}^{{select }}} M_p^{ {final }}(i, j)$                (12)

Pixels not covered by any activated category are labeled as background.

A small number of isolated noisy regions may still exist in the fused segmentation results, typically manifesting as small areas lacking consistency with surrounding semantic categories. A fully connected Conditional Random Field (CRF) is adopted as a post-processing step to eliminate such anomalous predictions. The CRF energy function consists of unary and pairwise potentials:

$E(y)=\sum_i \psi_u\left(y_i\right)+\sum_{i<j} \psi_p\left(y_i, y_j\right)$                  (13)

where, the unary potential $\psi_u\left(y_i\right)=-\log M^{ {final_{yi} }}(i)$ is directly taken from the fusion confidence map, and the pairwise potential $\psi_p\left(y_i, y_j\right)=\mu\left(y_i, y_j\right) \sum_{m=1}^K w_m k_m\left(f_i, f_j\right)$ imposes a penalty on label inconsistency between neighboring pixels based on color difference and spatial distance; $f_i$ is the combined feature of color and position of pixel $i$, and $k_m$ is a Gaussian kernel function. The CRF converges after five iterations, effectively eliminating isolated regions with an area smaller than 50 pixels, further improving segmentation results in terms of semantic continuity and boundary regularity.

2.4 Scene graph propagation and generative closed-loop feedback

Segmentation masks, as pixel-level dense prediction results, contain rich object instance information and spatial layout structures; however, they do not directly carry inter-object relations or high-order teaching connotations at the semantic level. Transforming segmentation results into a structured scene representation is a key bridge connecting pixel-level perception and scene-level reasoning. Figure 3 illustrates the integrated segmentation-reasoning-generation closed-loop feedback architecture. First, based on the segmentation output M, each connected component is treated as a node to construct a scene graph $G=(V, \mathrm{E})$. The node set $V=\left\{v_1, \ldots, v_M\right\}$ corresponds to all M segmentation regions, with an average of approximately 18.7 in experiments. The attributes of each node $v_k$ consist of three parts: semantic category $c_k$ taken from the ETST system; spatial bounding box $b_k=\left(x_k, y_k, w_k, h_k\right)$ describing the position and scale information of the region; and deep semantic features $f_k \in \mathrm{R}^{256}$ extracted from the second decoder layer of the U-Net as the visual representation. The edge set E encodes spatial topological relations between nodes, covering eight directional categories including containment, adjacency, left, right, above, and below. The spatial relation between nodes $v_i$ and $v_j$ is determined by comparing their bounding boxes:

$r_{i j}=\arg \max _{r \in \mathrm{R}} {Score}\left(r \mid b_i, b_j\right)$                (14)

where, R is the set of eight predefined spatial relations, and Score(⋅) is jointly computed based on the center coordinate offsets $\Delta x=x_i-x_j, \ \Delta y=y_i-y_j$, and the Intersection over Union $I o U\left(b_i, b_j\right)$, selecting the highest-scoring relation as the edge label.

3.png

Figure 3. Integrated segmentation-reasoning-generation closed-loop feedback architecture

After obtaining the scene graph, it needs to be transformed into a representation understandable by a LLM to complete teaching semantic reasoning. Directly constructing textual descriptions using category names and bounding box coordinates loses detailed visual feature information, and discrete category symbols poorly align with the continuous semantic space of language models. To this end, a graph-text alignment projection module is designed: a graph attention network is employed to aggregate neighborhood information of node features, and then a linear projection maps the graph structure into continuous soft prompt vectors. This process takes the form:

$S_{ {graph }}={Proj}\left({Aggregate}\left(\left\{f_k\right\}_{k=1}^M,\left\{r_{i j}\right\}_{i, j=1}^M\right)\right)$                 (15)

where, Aggregate is a two-layer graph attention network with a hidden dimension of 256; it takes node features $f_k$ as initial states and edge relations $r_{i j}$ as adjacency constraints, passing messages among nodes via an attention mechanism to enhance contextual awareness of node representations. Proj is a linear projection layer mapping the aggregated graph-level feature from 256 dimensions to 768 dimensions to align with the embedding space of the selected LLM. $S_{ {graph }}$ and the teaching reasoning prompt template $P_{{teach }}$ are jointly fed into the frozen LLaMA-2-7B model:

$H=L L a M A\left(S_{{graph }}, P_{ {teach }}\right)$              (16)

where, $P_{{teach }}$ contains instructions requiring the model to infer teaching information such as lesson topic, core knowledge points, and teacher-student interaction focus, and H is the model’s output structured teaching semantic description. During this process, all parameters of the LLM remain frozen, and only parameters of the graph attention network and projection layer participate in training. This design effectively controls the number of trainable parameters while retaining the strong reasoning ability of the large model.

The path from segmentation to scene understanding completes an ascending trajectory from pixels to semantics; however, such unidirectional propagation cannot allow errors at the segmentation stage to be perceived and corrected by subsequent modules. To establish a reverse optimization channel, a generative closed-loop feedback mechanism based on image reconstruction is introduced. The inferred teaching semantic description H and segmentation mask M are jointly injected as conditional signals into the decoder of Stable Diffusion to drive the reconstruction of an input image $I_{g e n}$. Specifically, H is converted into a conditional embedding via a fixed text encoder, and the downsampled segmentation mask M is concatenated with feature maps of each U-Net decoder layer along the spatial dimension, serving as a strong constraint on spatial structure. A joint reconstruction loss is constructed by comparing the difference between the reconstructed image and the original input:

$\begin{aligned} L_{ {cyc }}= & \left\|I-I_{\text {gen }}\right\|_1+\lambda_{ {perc }} \cdot\left\|\Phi(I)-\Phi\left(I_{ {gen }}\right)\right\|_2 +\lambda_{{ssim }} \cdot\left(1-\operatorname{SSIM}\left(I, I_{ {gen }}\right)\right)\end{aligned}$                 (17)

where, the first term is the pixel-wise L1 reconstruction error, ensuring basic pixel-level fidelity of reconstruction; $\Phi(\cdot)$ is a pre-trained Visual Geometry Group 16-layer (VGG-16) network, with activations of the relu3 3 layer taken as perceptual features, and the second term constrains the distance between reconstructed and original images in perceptual feature space to maintain visual perceptual consistency; the third term adopts the Structural Similarity Index (SSIM) to measure similarity in local contrast and structural information. Weight coefficients $\lambda_{{perc }}$ and $\lambda_{\text {ssim}}$ are set to 0.1 and 0.05 , respectively.

The core value of this reconstruction loss does not lie in generating realistic teaching images per se, but in its indirect supervision capability over segmentation quality. Its error back-propagation path is:

$\frac{\partial L_{c y c}}{\partial \Theta_{{gate }}} \approx \frac{\partial L_{c y c}}{\partial I_{ {gen }}} \cdot \frac{\partial I_{{gen }}}{\partial M_{{fused }}} \cdot \frac{\partial M_{{fused }}}{\partial \Theta_{ {gate }}}$                 (18)

where, $\Theta_{ {gate }}$ is the parameter set of the DAG network, and $M_{{fused }}$ is the fused attention map. The computation flow of this link is as follows: the gradient of $L_{ {cyc }}$ with respect to $I_{{gen }}$ is approximately solved through the cross-attention layers of the generative U-Net, reflecting how sensitively the generated image responds to input conditions; this gradient is then passed to the fused attention map $M_{{fused }}$ via the dependency of the generator on the segmentation mask, and finally back-tracks through the forward computational graph of the DAG network to the learnable parameters of fusion weights . A key design of this mechanism is that gradients only update parameters of the DAG network and prompt adapter, while backbone parameters of Stable Diffusion remain frozen. This means that although the reconstruction error originates from the generation process, its optimization objective is solely to adjust the multi-scale attention fusion manner at the segmentation stage, so that segmentation masks better serve subsequent conditional reconstruction. If the reconstruction quality of a local region in is markedly lower than surrounding areas, it often implies bias in the spatial constraint provided by the segmentation mask in that region; the error signal guides the DAG network to adjust the relative weights of self-attention across layers at that position, yielding more accurate boundary localization in the next iteration.

The joint optimization objective of the entire GSP-Net is a weighted combination of multiple losses:

$L_{{total }}=L_{C E}\left(M, M_{G T}\right)+\eta \cdot L_{c y c}+\lambda \cdot \sum_{l=1}^L\left\|\alpha_l-\alpha_l^{{init }}\right\|_2+\mu \cdot L_{ {rank }}$               (19)

The first term is the cross-entropy segmentation supervision loss, which is active only when labeled data is available; the second term is the closed-loop reconstruction loss, with the weight coefficient $\eta$ set to 0.5; the third term is a regularization term on attention weights, introduced to constrain the fusion weights $\alpha_l$ from undergoing excessive semantic drift during feedback iterations, with $\lambda$ set to 0.01 and $\alpha_l^{ {init }}=1 / L$ as the initial uniform weight; the fourth term is the hierarchical ranking loss, defined as:

$L_{ {rank }}=\sum_{l=1}^{L-1} \max \left(0, \alpha_{l+1}-\alpha_l+\delta\right)$               (20)

where, $\delta=0.05$ is a margin parameter. This loss encourages fusion weights of deeper decoders to be slightly higher than shallower ones; the rationale is that deep self-attention maps possess larger receptive fields and capture overall semantic object structures, thus should dominate fusion, while shallow self-attention maps play a supplementary role only in regions rich in local detail. This ranking constraint maintains a reasonable weight disparity between the two hierarchy types, avoiding deep features from being suppressed by shallow features in extreme cases. In fully annotation-free real teaching scenarios, weights are set as $\eta=1.0, \lambda=0.1$, and the $L_{C E}$ term is deactivated; at this point, the closed-loop reconstruction loss acts as the primary supervision signal, driving the segmentation module toward self-supervised evolution through continuous interaction in teaching scenes.