English Scene Character Detection via Multi-Scale Contextual Feature Fusion and Attention Mechanisms

2.1 Overview of the overall detection framework

An end-to-end deep learning framework for character detection in complex English scenes is proposed in this work, with the overall network following a progressively optimized feature processing logic to accomplish character detection tasks. The model takes raw scene images as input, whereupon basic multi-scale feature extraction is first performed via a backbone network. ResNet-50 is selected as the default backbone, while the Swin Transformer architecture is also accommodated to ensure feature modeling flexibility. Deep features at three different resolutions are extracted from the backbone network as fundamental representations, with each level of features corresponding to a distinct downsampling scale, thereby effectively covering multi-dimensional feature information of scene characters and balancing shallow spatial detail features with deep semantic features, thus providing stable foundational inputs for subsequent fine-grained feature optimization and detection inference.

The overall network architecture is composed of four core functional modules that are sequentially arranged to form a complete inference pipeline, with each module being orderly connected according to the logic progressing from coarse-grained feature extraction to fine-grained result optimization. The foundational features are first fed into the multi-scale contextual perception feature extraction module, where multi-scale semantic mining and long-range contextual dependency modeling are accomplished, compensating for the scale-adaptability deficiencies inherent in conventional feature representations. Subsequently, collaborative enhancement and dynamic fusion across feature dimensions are realized through the dual-path attention adaptive fusion module, thereby improving the specificity and effectiveness of feature representations. The model further introduces a dynamic upsampling mechanism for feature map resolution reconstruction, whereby contour distortions introduced by conventional sampling methods are avoided while feature dimensions are restored. Finally, iterative optimization of instance contours and morphological constraint correction are achieved through the character-level progressive boundary refinement strategy, with high-accuracy scene English character detection results being output. Each module provides complementary optimization in a progressive manner, enabling simultaneous enhancement of both detection accuracy and robustness for English characters in complex scenes.

2.2 Multi-scale contextual perception feature extraction module

To address the challenges posed by the large-scale span of English characters in natural scenes and the insufficient receptive field coverage of conventional feature extraction structures, a multi-scale contextual perception feature extraction module is designed in this work, whereby multi-dimensional semantic features and long-range contextual information are simultaneously captured within a controllable computational budget. The architecture of the multi-scale contextual perception feature extraction module is illustrated in Figure 1. The module takes the multi-level features output by the backbone network as input, whereupon a 1 × 1 convolutional operator is first applied to compress the feature channels, reducing the original channel dimension to one-quarter and thereby effectively decreasing the parameter count and computational complexity of subsequent convolutional operations. The compressed features are then fed into three parallel 3 × 3 atrous convolutional branches, through which receptive field representations of varying extents are obtained via the configuration of differentiated dilation rates, thus accommodating the feature extraction requirements of both extremely small-scale and extremely large-scale characters in scenes. The dilation rates of each branch are sequentially set to 3, 6, and 12 in this work, ensuring gradient coverage of the receptive field from local regions to large-scale regions, with each branch maintaining uniform input and output channel dimensions to guarantee multi-scale feature dimensional alignment. The branch output features can be computed as follows:

$F^{(r)}=\operatorname{Conv}_{3 \times 3}^{(r)}\left(F_{\text {down}}\right), r \in\{3,6,12\}$   (1)

where, $F_{\text {down}}$ denotes the channel-compressed foundational features; $\operatorname{Conv}^{(r)_{3 \times 3}}$ represents the atrous convolution operation with dilation rate $r$; and $F^{(r)}$ denotes the multi-scale feature output corresponding to a single convolutional branch.

image005.png

Figure 1. Architecture of the multi-scale contextual perception feature extraction module

To compensate for the deficiency of multi-scale atrous convolutions in modeling pixel-level long-range dependencies, a criss-cross attention mechanism is introduced in this work to perform semantic enhancement on the features of each branch, focusing on pixel-wise associative features along the horizontal and vertical orthogonal directions of the image, which is well-aligned with the horizontally arranged spatial distribution characteristics of English characters. For the output features of each atrous convolution branch, query, key, and value feature maps are generated through linear transformations, with the feature channel dimension being uniformly compressed to one-eighth of the input dimension, thereby preserving core semantic information while reducing the computational cost of attention operations. The attention weights are computed independently along the row and column dimensions of the image, and contextual features from the entire global row and column regions are aggregated pixel-wise, enabling effective modeling of long-range semantic associations within character regions. The feature enhancement process can be defined as:

$y_{i, j}=\sum_{k=1}^W \operatorname{Softmax}\left(q_{i, j}^{\top} k_{i, k}\right) v_{i, k}+\sum_{l=1}^H \operatorname{Softmax}\left(q_{i, j}^{\top} k_{l, j}\right) v_{l, j}$   (2)

where, $H$ and $W$ denote the height and width of the feature map, respectively; $q_{i, j}, k_{i, k}$, and $v_{i, k}$ denote the query vector, key vector, and value vector corresponding to spatial position $(i, j)$ of the feature map; and $y_{i, j}$ denotes the pixel-wise feature after attention enhancement. Compared with the quadratic complexity of global selfattention, this mechanism reduces the overall computational complexity from $O\left(H^2 W^2\right)$ to $O(H W(H+W))$, achieving a reduction of over 85% in computational cost for input images at $1280 \times 1280$ resolution, thereby realizing an effective balance between high-precision contextual modeling and low computational overhead.

After the completion of single-branch multi-scale feature enhancement, the feature information from the three branches with differentiated receptive fields is deeply fused. An element-wise additive strategy is adopted in this work for preliminary aggregation of the multi-branch features, whereby fine-grained texture information and long-range contextual semantics captured by each branch are completely preserved through this fusion approach, while feature redundancy and information offset issues caused by channel concatenation operations are avoided. Cross-channel information interaction and dimension restoration are then performed on the preliminarily aggregated features via a 1 × 1 convolution, with the feature channels being restored to the original dimensions of the module input, thereby achieving complementary coupling and unified representation of multi-scale semantic features. Following the multi-stage processing, the module ultimately outputs the optimized multi-level enhanced features $\widetilde{F}_3, \widetilde{F}_4$, and $\widetilde{F}_5$, which effectively compensate for the deficiencies of conventional feature pyramids in extreme-scale character representation and long-range contextual dependency modeling, providing highly robust foundational feature support for subsequent feature fusion and fine-grained detection tasks.

2.3 Dual-path attention feature enhancement and adaptive fusion module

To address the issues of channel semantic redundancy and weakened spatial localization information inherent in multi-scale contextual features, a dual-path attention feature enhancement and adaptive fusion module is constructed in this work, whereby the quality of feature representations is simultaneously optimized from two dimensions—channel semantic modeling and spatial position reinforcement—enabling fine-grained calibration and scene-adaptive fusion of multi-scale features. The architecture of the dual-path attention feature enhancement and adaptive fusion module is illustrated in Figure 2. The module first performs channel-wise feature selection and dependency modeling through a grouped Transformer structure, with global semantic statistics being computed for the enhanced features at each level. For a hierarchical feature $\widetilde{F}_l$, global spatial information is aggregated via global average pooling, generating a one-dimensional channel descriptor that characterizes the overall response intensity of each channel, which is computed as follows:

$z_l=\frac{1}{H_l W_l} \sum_{i=1}^{H_l} \sum_{j=1}^{W_l} \widetilde{F}_l(i, j), z_l \in \mathrm{R}^{C_l}$   (3)

where, $H_l$ and $W_l$ denote the spatial dimensions of the $l$-th level feature map, and $C_l$ denotes the channel dimension of the feature. To balance channel interaction modeling capacity with computational efficiency, the channel descriptor is uniformly divided into eight sub-feature groups, within each of which inter-channel relationships are mined through a lightweight Transformer composed of a four-head multi-head self-attention mechanism combined with a two-layer perceptron. After the completion of feature interactions within each group, dimensional concatenation is performed, and normalized channel weight coefficients are generated via a Sigmoid activation function. Finally, semantic calibration of the original features is accomplished through element-wise multiplication, whereby effective feature channels are strengthened and ineffective noise channels are suppressed. The calibration process can be expressed as:

$\widehat{F}_I=\alpha_I \odot \widetilde{F}_I$   (4)

where, $\alpha_l$ denotes the normalized channel attention weights, ⊙ denotes the element-wise multiplication operation, and $\widehat{F}_l$ denotes the features after channel-dimensional optimization.

image017.png

Figure 2. Architecture of the dual-path attention feature enhancement and adaptive fusion module

Building upon the channel semantic calibration, a bidirectional coordinate attention mechanism is introduced into the module to achieve spatial-dimensional position information enhancement, whereby character target region features are precisely activated and localization deviations for small-scale and narrow characters are ameliorated. Global pooling sampling is performed on the calibrated features along two orthogonal directions, namely horizontal and vertical, through this mechanism, generating bidirectional spatial descriptors containing precise positional information, thereby effectively overcoming the limitation of conventional spatial attention in encoding long-range positional associations. The bidirectional pooling operations are defined as:

$p_l^h(i)=\frac{1}{W_l} \sum_{j=1}^{W_l} \widehat{F}_l\left(i_j, j\right), p_l^h \in \mathrm{R}^{C_l \times H_l \times 1}$   (5)

$p_l^v(j)=\frac{1}{H_l} \sum_{i=1}^{H_l} \hat{F}_l(i, j), p_l^v \in R^{C_l \times 1 \times W_l}$   (6)

where, $p_l^h$ and $p_l^v$ denote the positional feature descriptors along the horizontal and vertical directions, respectively. The two types of one-dimensional positional features are expanded to a unified spatial size via dimensional broadcasting, after which channel concatenation is performed. A 2D spatial attention weight matrix is then generated through convolution for dimensionality reduction, nonlinear activation, and dimensionality restoration. The weight matrix is applied to the input features via element-wise weighting, achieving feature-adaptive enhancement in the image spatial dimension, with the spatially optimized features being ultimately obtained as:

$\breve{F}_l=\beta, \odot \widehat{F}_l$   (7)

where,$\beta_l$ denotes the normalized spatial attention weights. To address the limitation that fixed-weight fusion strategies cannot adapt to the feature distributions of complex scenes, an adaptive gating fusion mechanism is designed in this work, whereby the fusion proportions between channel semantic features and spatial positional features are dynamically balanced. Global semantic statistics are performed on the spatially enhanced features through this mechanism, with spatial dimensions being compressed via global average pooling, feature mapping and dimensionality transformation being accomplished through a two-layer fully connected network, and an adaptive gating coefficient in the range of 0 to 1 being ultimately output via a Sigmoid function. The computation process is defined as follows:

$\gamma_l=\operatorname{Sigmoid}\left(F C\left(G A P\left(\check{F}_l\right)\right)\right)$   (8)

where, $\gamma_l$ denotes the adaptive fusion weight coefficient, and the intermediate fully connected layer dimension is set to one-sixteenth of the original channel dimension, thereby reducing computational overhead while maintaining mapping capacity. Based on this dynamic weight, the features from the channel-enhanced branch and the spatially enhanced branch are fused in a weighted manner, with the fusion formula given as:

$F_l^{\text {fuse}}=\gamma_l \cdot \check{F}_l^{\text {chan}}+\left(1-\gamma_l\right) \cdot \check{F}_l^{\text {spat }}$   (9)

Through this fusion approach, the contribution proportions of the two-branch features are automatically adjusted according to the scene complexity, character scale, and background interference level of the input image, with fine spatial details being preferentially preserved in simple scenes and high-level semantic features being emphasized in complex interference scenes. High-quality multi-scale fused features that combine both semantic representation capacity and spatial localization precision are ultimately generated, providing superior feature support for subsequent upsampling reconstruction and boundary detection tasks.

2.4 Dynamic upsampling module with multi-scale context enhancement

During the feature decoding stage of scene character detection networks, conventional upsampling operations commonly rely on transposed convolutions for feature map resolution restoration, with fixed-parameter convolutional kernels inevitably producing periodic grid overlap effects that induce checkerboard artifacts. While such distortion issues have limited impact in generic object detection tasks, they are extremely sensitive to English narrow characters with minimal pixel widths, readily disrupting the continuity and integrity of character edges, causing feature contour distortion, and ultimately resulting in detection boundary regression deviations. To address the detail loss and contour distortion issues caused by fixed sampling patterns, a dynamic upsampling module with multi-scale context enhancement is designed in this work, whereby sampling coordinates are adaptively corrected through the neighborhood semantic information of the features themselves, freeing the operation from the constraints of fixed convolutional kernels. High-fidelity feature reconstruction with both enhanced resolution and complete preservation of character edge details is thereby achieved. A schematic illustration of the context-driven dynamic upsampling process is shown in Figure 3.

image035.png

Figure 3. Schematic illustration of the context-driven dynamic upsampling process

The correction accuracy of dynamic sampling coordinates is highly dependent on the quality of feature information in the neighborhood of each sampling point; therefore, multi-scale contextual feature extraction is first performed on each interpolation reference position within the module. With the network upsampling rate being denoted as s, the reference sampling coordinates on the source feature map corresponding to spatial coordinates $\left(i^{\prime}, j^{\prime}\right)$ on the target feature map can be mapped as $(i, j)=\left(i^{\prime} / s, j^{\prime} / s\right)$. The neighborhood sampling range is delineated centered at the reference coordinates, through which multi-scale local feature information is mined in parallel via three groups of atrous convolutions with different dilation rates, capturing fine edge textures and large-range neighborhood association features simultaneously. The aggregated neighborhood contextual feature set is defined as:

$N_{i, j}=\bigcup_{r \in\{1,2,4\}}\left\{\operatorname{Conv}_{3 \times 3}^{(r)}\left(F^{\text {fuse}}\right)_{i+\Delta i, j+\Delta j}:|\Delta i|,|\Delta j| \leq k / 2\right\}$   (10)

where, $r$ denotes the atrous convolution dilation rate, $k$ denotes the neighborhood window size, $F^{\text {fuse}}$ denotes the fused features output by the preceding dual-path attention module, and $\Delta i$ and $\Delta j$ denote the neighborhood spatial offsets. The multi-group scale features are concatenated and fused along the channel dimension, constructing a feature vector that contains multi-level local semantics and edge information, thereby providing sufficient and fine-grained feature support for subsequent coordinate offset prediction, while sampling deviations caused by insufficient single-scale contextual information are avoided.

Based on the aggregated neighborhood contextual features, a lightweight convolutional network is employed within the module to perform dynamic offset prediction of sampling coordinates. The prediction network is composed of two stacked 1 × 1 convolutional layers, with the first layer reducing the input feature channel dimension to 64 to reduce computational cost, and the second layer outputting two-dimensional coordinate offset parameters corresponding to horizontal and vertical position corrections, respectively. The prediction process can be expressed as:

$\left(\Delta i^*, \Delta j^*\right)=\Phi\left(N_{i, j}\right)$   (11)

where, $\Phi$ denotes the mapping function of the offset prediction network, and $\Delta i^*$ and $\Delta j^*$ denote the final optimized coordinate offsets. To suppress feature disorganization caused by excessive coordinate offsets during training and to enhance model convergence stability, the offset values are constrained within the range of ±1.5 pixels. Based on the corrected precise sampling coordinates, bilinear interpolation is performed on the original fused feature map within the model, enabling adaptive computation of pixel values at the target positions, with sampling positions being dynamically adapted according to local semantics, thereby fundamentally avoiding the regularized artifacts and edge distortions introduced by fixed sampling patterns.

To accommodate the multi-level feature fusion architecture, differential upsampling ratios are applied to the fused features at three different scales. Upsampling ratios of 2×, 4×, and 8× are respectively applied to F^fuse₃, F^fuse₄, and F^fuse₅, with all level features being uniformly mapped to a resolution scale of H/4 × W/4. The multi-level features after size unification are deeply fused via channel concatenation, producing a high-fidelity feature representation that combines both shallow spatial details and deep semantic information, serving as the input for the subsequent character boundary refinement module. This sampling mechanism is fully adapted to the feature distribution characteristics of English narrow characters and small-scale characters, effectively compensating for the detail deficiencies of conventional upsampling methods, and providing a high-quality feature foundation for subsequent high-precision character boundary regression and contour fitting.

2.5 Character-level progressive boundary refinement strategy

To achieve sub-pixel-level boundary fitting and morphological normalization correction for English characters, a coarse-to-fine character-level progressive boundary refinement strategy is designed in this work, whereby detection contour deviations are progressively corrected through a multi-level iterative optimization logic. Coarse-grained character instance prediction is first performed within the model, providing an initial contour basis for subsequent fine-grained optimization. A character region probability map of dimension $H / 4 \times W / 4$ is output through the segmentation branch within the network, with a precise character foreground mask being generated in conjunction with a differentiable binarization algorithm, thereby effectively distinguishing target regions from complex backgrounds. Simultaneously, a four-channel offset parameter map is output through the boundary regression branch, with pixel-wise distances from the current position to the top, bottom, left, and right boundaries of each character being predicted. Instance partitioning of the prediction results is accomplished through connected component analysis, from which the initial bounding box and complete contour curve of each individual character are obtained, completing the localization output of the coarse detection stage. An iterative flow diagram of the character-level progressive boundary refinement strategy is shown in Figure 4.

To address the issues of feature redundancy and detail deficiency inherent in conventional uniform sampling approaches, a curvature-adaptive keypoint sampling mechanism is adopted in this work for boundary point selection. The initial character contour is constructed as a continuous parametric curve $C(t)=(x(t), y(t))$, with the parameter variable t being normalized to the range of 0 to 1. The curvature magnitude $\left\|C^{\prime \prime}(t)\right\|$ at each position is obtained by solving the second-order differential of the curve, quantitatively characterizing the local bending degree of the contour. Adaptive sampling point selection is achieved based on the cumulative curvature distribution characteristics, with sampling being densified in regions of severe contour deformation and sparsified in smooth regions. The sampling point selection rule is defined as:

$t_k=\inf \left\{t: \int_0^t\left\|C^{\prime \prime}(\tau)\right\| d \tau \geq \frac{k}{N} \int_0^1\left\|C^{\prime \prime}(\tau)\right\| d \tau\right\}, k=1,2, \ldots, N$   (12)

where, t_k denotes the curve parameter corresponding to the k-th sampling point, and N denotes the total number of sampling points per character contour, which is uniformly set to 16 in this work. Through this sampling approach, the geometric features of character corners and curved edges are maximally preserved without increasing computational overhead, thereby enhancing the fineness of contour representation.

image050.png

Figure 4. Iterative flow diagram of the character-level progressive boundary refinement strategy

After the completion of boundary keypoint sampling, iterative boundary position optimization is performed within the model based on multi-scale neighborhood contextual features. In each iteration of the update process, dual-layer neighborhood features with radii of 3 and 7 are extracted centered at each boundary keypoint, fusing local fine-grained textures with large-range contextual semantics to compensate for the limitations of single-point feature representation. The multi-layer neighborhood features are concatenated with the original keypoint features and fed into a lightweight regression network, which is composed of two convolutional layers and two fully connected layers in series, for predicting the spatial position residuals of the keypoints. The computation is defined as follows:

$\Delta p_i^{(t)}=\Psi\left(\operatorname{Concat}\left(F e a t\left(p_i\right), C o n t e x t\left(p_i, r_1\right), C o n t e x t\left(p_i, r_2\right)\right)\right)$   (13)

where, $\Psi$ denotes the mapping function of the boundary regression network, and $\Delta p_i^{(t)}$ denotes the position correction residual at the $t$-th iteration. Based on the residual results, the coordinate positions are updated point-by-point as $p_i^{(t+1)}=p_i^{(t)}+\Delta p_i^{(t)}$, with high-precision convergence of boundary coordinates being achieved after three iterations of optimization, thereby effectively eliminating localization deviations in the coarse contours.

To address the issue of character morphological distortion that tends to occur during boundary optimization in complex scenes, the inherent aspect ratio prior of English characters is introduced as a regularization constraint in this work, thereby enhancing the physical plausibility of the detection results. In accordance with the morphological distribution characteristics of English characters, the detected targets are categorized into three classes—uppercase letters, common lowercase letters, and narrow-bodied letters—with corresponding standard aspect ratio reference ranges being established. During the training phase, the reference parameters are matched according to the ground-truth categories, while during the inference phase, category adaptation is automatically accomplished through the dimensional characteristics of the coarse segmentation contours. A regularization loss function is constructed by computing the error between the aspect ratio of the predicted contour bounding rectangle and the standard reference value:

$L_{\text {prior}}=\sum_i\left\|A R\left(p_i^{(t+1)}\right)-A R_{\text {ref}}\right\|_2^2$   (14)

where, $A R\left(p_i^{(t+1)}\right)$ denotes the actual aspect ratio of the optimized character contour, and AR_ref denotes the standard reference value corresponding to the character category. This constraint mechanism effectively suppresses abnormal morphological deviations during the boundary regression process, adapts to the structural characteristics of narrow English characters, and ultimately achieves character boundary detection results that combine both high precision and morphological plausibility.

2.6 Overall loss function and end-to-end training strategy

A composite loss function is constructed in this work based on the principle of multi-task collaborative optimization, through which character region segmentation, coarse-grained boundary regression, fine-grained boundary fitting, and character morphological constraints are jointly supervised, enabling end-to-end parameter optimization of the entire network framework. The overall loss is obtained through the weighted combination of four functionally complementary sub-losses, with the specific expression given as:

$L_{\text {total}}=\lambda_1 L_{\text {seg}}+\lambda_2 L_{\text {reg}}+\lambda_3 L_{\text {boundary}}+\lambda_4 L_{\text {prior}}$   (15)

where, $\lambda_1$, $\lambda_2$, $\lambda_3$, and $\lambda_4$ denote the balancing weights for each loss term, with the optimal weight configuration of $1.0,1.0,0.8$, and 0.2 being determined through grid search traversal of the parameter space in this work. The segmentation loss $L_{\text {seg}}$ is trained with an equalweight combination of dice loss and focal loss, whereby the sample imbalance issue between foreground characters and background regions in natural scenes is alleviated, while the model's learning capacity for hard samples such as blurred, occluded, and extremely small characters is simultaneously enhanced. The coarse boundary regression loss $L_{\text {reg}}$ adopts Smooth L1 loss to supervise pixel-level boundary offset parameters, ensuring stable convergence of the initial detection box localization. The boundary refinement loss $L_{\text {boundary}}$ measures the deviation between the iteratively optimized keypoints and the ground-truth contours using L1 distance, precisely constraining sub-pixel-level boundary fitting performance. The morphological prior loss $L_{\text {prior}}$ incorporates the distribution characteristics of English character aspect ratios to construct a regularization constraint, effectively correcting morphological distortions that occur during the boundary optimization process, thereby improving the plausibility and authenticity of the detection results.

To balance model training stability with scene generalization capacity, a two-stage progressive training strategy is adopted in this work for iterative network parameter updates. The model is first pre-trained on the large-scale SynthText synthetic dataset for 100 epochs, through which general texture features and spatial arrangement regularities of English characters are comprehensively learned, a stable foundational feature representation capacity is established, and the overfitting risk caused by the limited sample size of real-scene datasets is effectively avoided. After the completion of the pre-training phase, targeted fine-tuning is performed on the Total-Text and ICDAR 2015 real-scene datasets, adapting to real-scene characteristics such as complex background interference, severe scale variations, and curved characters. A learning rate warm-up mechanism is introduced during the training process, whereby the learning rate is gradually increased in the initial iterations, thereby avoiding convergence instability caused by parameter update oscillations. Concurrently, a gradient clipping strategy is introduced to constrain the range of extreme gradient values, with the gradient explosion problem during deep network training being suppressed, thereby ensuring stable parameter updates throughout the multi-module collaborative training process and fully unleashing the feature optimization and boundary refinement performance of each core module.

The performance advantages of the proposed detection framework are derived from the progressive collaborative design and scene-adaptive optimization of each core module, with different functional modules addressing the inherent challenges of English scene character detection from the perspectives of feature extraction, feature fusion, image reconstruction, and boundary optimization. Through the construction of full-coverage receptive fields via gradient-dilated convolutions, coupled with lightweight long-range dependency modeling through criss-cross attention, the multi-scale contextual perception feature extraction module effectively resolves the insufficiency of conventional networks in capturing extremely large- and small-scale character features, thereby enhancing multi-scale scene adaptability at the foundational feature level. The dual-path attention adaptive fusion mechanism performs channel-wise semantic screening and spatial position enhancement separately, with adaptive feature fusion across different scenes being achieved through dynamic gating weights, thereby overcoming the limitations of fixed fusion strategies while balancing noise suppression in complex backgrounds with fine character detail preservation. By abandoning the inherent deficiencies of fixed-kernel sampling and adaptively correcting sampling coordinates based on neighborhood context, the dynamic upsampling mechanism thoroughly alleviates the destructive effects of checkerboard artifacts on narrow character contours. Through curvature-adaptive sampling for enhanced contour representation accuracy and aspect ratio prior constraints for reduced boundary regression solution space, the character-level boundary refinement strategy effectively avoids morphological distortions caused by unconstrained optimization, achieving simultaneous improvements in both detection accuracy and contour plausibility.

Although excellent detection performance is exhibited by the proposed method in conventional complex natural scenes, significant limitations remain under extreme operating conditions and engineering deployment scenarios. In highly occluded scenes, when the effective character area occlusion ratio exceeds 50%, substantial loss of local character texture and neighborhood contextual information occurs, and the boundary iterative optimization module fails to obtain sufficient feature support for coordinate correction, resulting in notable detection accuracy degradation. For non-standardized characters such as artistically deformed fonts and irregular handwriting, the morphological scales deviate from the aspect ratio distribution patterns of conventional printed fonts, rendering the preset prior constraints inapplicable to such irregular character structures and substantially reducing their effectiveness. At the model deployment level, although steady accuracy improvements are achieved through the multi-module collaborative architecture, the parameter count remains relatively large compared with lightweight detection models, making it difficult to directly adapt to real-time detection requirements of low-computational-power terminal devices such as embedded and mobile platforms, thereby limiting the practical application scenarios of the algorithm.

To address the limitations of the current method, three future research directions are identified in this work: robustness optimization, lightweight deployment, and generalization expansion. In subsequent research, high-level textual semantic priors may be introduced, leveraging character sequence distribution regularities and English lexical knowledge to assist feature inference, thereby compensating for the deficiency of low-level image features in extreme occlusion and blurring scenarios and further enhancing the environmental robustness of the model. Concurrently, structured pruning and knowledge distillation techniques can be incorporated to streamline redundant network parameters and feature channels, with feature knowledge transfer being accomplished through high-precision teacher models, thereby reducing model size and computational cost while maintaining detection accuracy, and adapting to edge-device deployment requirements. Furthermore, generalization-oriented modifications can be performed on the existing network architecture to adapt to the morphological characteristics and structural regularities of multilingual characters, transcending the single-language English detection constraint and constructing a universal scene character detection framework applicable to multiple languages and fonts, thereby further expanding the academic value and engineering application scope of the algorithm.