Female Identity and Class Imagination in Qing Chinese Export Paintings: An Artificial Intelligence-Assisted Study of Cross-Cultural Interpretation

Summary

This article presents a human-supervised computational workflow to examine recurrent representations of women in 433 Qing-era Chinese export paintings. Seven human-reference categories and recurring visual cues characterize patterns within the corpus, while the findings remain bounded to the sampled images and do not measure Qing social reality or historical reception.

Abstract

This study examined recurring representations of female identity and class-coded visual cues in a corpus of 433 Qing Chinese export paintings dated 1757–1911. The workflow combined CLIP ViT-B/32 for preliminary labeling, Segment Anything Model (SAM)-assisted localization, dimensionality reduction and clustering, and manual iconographic coding by two independent coders. CLIP labels were used for preliminary organization, SAM served as a localization aid, and adjudicated human coding provided the reference categories. The aggregate results identified seven human-reference categories: elite/beauty (n = 112), laboring (n = 72), domestic (n = 64), servant (n = 58), ritual/wedding (n = 45), market (n = 43), and performing (n = 39). Preliminary computational labels matched the human-reference labels for 356 of 433 paintings (82.2%), with category-level recall ranging from 73.3% to 88.4%. Recurring combinations of clothing, spatial setting, objects, posture, and depicted social relations were used to characterize the visual profiles of these categories. Evaluation of the 433 case-level AI–human label pairs yielded comprehensive category-level precision, recall, and F1 scores, alongside a confusion matrix detailing classification patterns. Cluster validation supported the selected clustering solution and demonstrated correspondence between the derived clusters and human-defined reference categories. Overall, recurrent combinations of clothing, spatial setting, objects, posture, and social relations indicate patterned pictorial typologies within the sampled corpus. These findings characterize visual representation in Qing Chinese export paintings but should not be interpreted as direct measures of legal status, lived experience, market demand, or historical foreign reception.

Introduction

Qing Chinese export paintings formed part of the commercial and cultural networks linking Canton and other collecting centers with Euro-American buyers from the late eighteenth century onward^1,2,3,4,5. Merchants, travelers, diplomats, missionaries, and collectors acquired these paintings as portable commodities, souvenirs, sources of visual information, and representations of places, occupations, production, customs, and social types. Their circulation, therefore, reflected both Chinese workshop practices and the expectations of external markets.

Foreign demand favored recognizable subjects, serial formats, repeated motifs, and scenes that could be collected and compared. Workshops adapted media, scale, composition, and subject matter for sale, while buyers often favored legible depictions of trades, ceremonies, interiors, gardens, streets, and picturesque social types. Repeated versions of Canton workshop scenes further demonstrate how artists and assistants adapted iconography and Western pictorial devices for overseas patrons^1,2,3,4,5. Although market demand may have influenced what was produced and preserved, the present corpus does not directly measure individual buyers’ preferences or responses.

These paintings should therefore be regarded as selective and mediated sources rather than comprehensive records of everyday life. Their scenes may idealize, simplify, standardize, or adapt social practices through workshop conventions, market selection, museum acquisition, cataloging, and digitization^1,2,3,4,5. Historical interpretation must distinguish features directly observable in the paintings from inferences about the individuals represented, the conditions of production, and the meanings later viewers attributed to the works.

Female figures provide a focused means of examining this distinction. Clothing, hairstyle, posture, spatial setting, objects, activities, and relationships with other figures can organize women into visually recurring roles, including elite/beauty, domestic, servant, laboring, market, performing, and ritual/wedding categories. These categories are analytical descriptions of pictorial patterns and do not, by themselves, establish legal rank, wealth, occupation, ethnicity, Banner or Han affiliation, moral character, or lived identity.

Digital art history and computer vision can facilitate large-scale comparisons of digitized visual collections while also enabling critical examination of the assumptions embedded in computational models^6,7. Vision–language models can assist in identifying and organizing recurring visual motifs, although cultural bias remains an important limitation when contemporary model vocabularies are applied to historical, non-Western images⁸. Related computational research on traditional Chinese painting similarly demonstrates the value of combining quantitative pattern detection with historically informed interpretation⁹.

This study addresses three questions: (1) Which adjudicated human-coded female identity categories are represented in the corpus? (2) Which combinations of clothing, spatial setting, objects, posture, and social relations characterize these categories? (3) How do overall exact-match accuracy and category-level recall vary across the seven categories? The workflow draws on computational approaches to traditional Chinese painting⁹, while restricting its historical claims to the sampled images and the documented analyses.

The intended contribution is both historiographical and technical. Corpus-scale comparison can identify recurrent pictorial formulas that can subsequently be contextualized through close historical analysis. This approach can inform research on gender, hierarchy, representations of everyday life, and cross-cultural encounters without treating image frequency as evidence of social prevalence or visual similarity as proof of historical identity.

Protocol

Study design and analytical unit
Each painting or catalog record was treated as one analytical unit. A corpus-based observational design was used, combining metadata review, preliminary computational organization, independent manual coding, and art-historical interpretation. The study period was defined as 1757–1911, and inclusion and exclusion criteria were applied within this timeframe. Historical artworks and their associated catalog metadata were analyzed. When distributions across date intervals of unequal duration were compared, counts were normalized by interval duration, while potential survival and acquisition biases were considered in the interpretation.

Corpus construction and selection
Candidate records were retrieved from the cited museum catalogs, digital archives, and published catalogs. For each candidate record, the institution, title, date or date range, medium, accession or catalog number, stable source URL, retrieval date, and rights statement were retained. Source links were preserved even when the corresponding image file could not legally be redistributed. The reported screening procedure began with 612 candidate images. Of these, 85 were excluded because they fell outside the 1757–1911 period or the export-painting tradition, 58 were excluded because they were modern reproductions, blurred records, or depicted incorrect subjects, 29 were excluded because they lacked an identifiable female figure or adequate image resolution, and 7 were excluded because of insufficient metadata. The final analytic corpus comprised 433 records.

Metadata standardization and image preprocessing
Each painting in the final analytic corpus was assigned a stable QEP identifier ranging from QEP_001 to QEP_433. Associated museum metadata were retained, including institution, accession number, title, date, medium or materials, place of origin, source link, and available catalog description. For visual analysis, the preliminary AI label, adjudicated human category, number of female figures, age group, posture, clothing, spatial setting, associated objects, and social relations were recorded. Each painting was classified into one of seven primary categories: elite/beauty, domestic, servant, laboring, market, performing, or ritual/wedding. For scenes containing multiple female figures, the focal figure was identified according to visual prominence and narrative centrality. When no single female figure was dominant, the primary category was assigned based on the principal depicted activity and the overall scene context. Overlapping categories were resolved by prioritizing depicted activity and scene context over clothing or appearance alone. Ambiguous cases were independently reviewed by both coders and resolved through consensus. Source images were saved in JPEG or PNG format. Only webpage borders, catalog frames, or non-image margins were cropped. Painted content was not reconstructed, recolored, enhanced, or restored.

CLIP-based preliminary labeling
The CLIP ViT-B/32 image encoder was used for preliminary image–text similarity analysis9. Candidate identity terms included elite woman, domestic woman, female servant, working woman, market woman, female performer, bride, and ritual woman. Scene terms included Chinese interior, garden, street market, workshop, tea production, textile work, and ceremonial scene. The complete set of prompt templates, their order, any prompt ensembling, text normalization procedures, decision rules, confidence scores, and ties were retained for each image.

The complete prompt list was applied consistently to all images using the OpenAI CLIP ViT-B/32 model implemented in Python 3.10 with PyTorch 2.0.1 and the OpenAI CLIP package. Inference was performed on a CUDA-enabled GPU with a batch size of 32 using FP32 precision. Each image was resized and center-cropped to 224 × 224 pixels, and the RGB channels were normalized using the preprocessing transformation associated with the ViT-B/32 model. Text prompts were constructed using the identity and scene descriptors defined above and encoded with the CLIP text encoder. Cosine similarity was calculated between normalized image and text embeddings, and the prompt with the highest similarity score was assigned as the preliminary AI label. Similarity scores for all candidate labels were retained for subsequent human review. A fixed random seed of 42 was used, and identical preprocessing procedures, prompt templates, and decision rules were applied to all 433 paintings.

SAM-assisted localization
The Segment Anything Model (SAM) was strictly utilized to localize figures, garments, furniture, tools, ritual objects, and architectural regions for human visual inspection. It must be emphasized that SAM did not participate in the classification pipeline; its masks, crops, and derived features were isolated from CLIP labeling and clustering processes, ensuring that SAM's application provided localization aids without influencing or inflating classification performance.

Manual point prompts or bounding boxes were applied when necessary to refine the localization of visual elements. SAM masks were generated for female figures, garments, furniture, tools, ritual objects, and architectural elements, and each segmentation result was visually inspected for appropriate regional coverage. The resulting masks supported the identification and review of relevant iconographic features but were not used for CLIP labeling or category assignment.

Dimensionality reduction and clustering
CLIP image embeddings were computed, and UMAP with cosine distance was used for two-dimensional visualization¹⁰. The representation used for clustering, together with its preprocessing procedure, dimensionality, and distance definition, was recorded.

K-means and hierarchical clustering were applied to the image-feature representations to identify recurrent visual groupings within the corpus¹¹. Candidate K-means solutions for k = 5–9 were evaluated using the silhouette coefficient and Davies–Bouldin index. K-means++ initialization was used with n_init = 10, max_iter = 300, tol = 1 × 10⁻⁴, and random_state = 42. Hierarchical clustering was performed using agglomerative clustering with average linkage and cosine distance. Candidate solutions were compared using quantitative validation indices, cluster stability, and art-historical interpretability, and the final cluster solution was selected for its most coherent representation of recurrent visual patterns. The resulting cluster assignment for each of the 433 paintings was retained for subsequent comparison with the human-coded categories.

Manual coding, training, and adjudication
Two coders independently reviewed all 433 images using a versioned codebook covering primary category, figure count, age group, posture, clothing, spatial setting, objects, and social relations. Coder-specific records were retained before adjudication. The coders had relevant backgrounds in art history and visual analysis and were trained using the standardized coding manual and representative examples from the painting corpus. Before formal coding, they completed a calibration exercise involving 30 paintings to promote consistent interpretation of the seven identity categories and five visual-cue dimensions.

During formal coding, both coders independently evaluated all eligible paintings while blinded to each other’s coding decisions, preliminary AI labels, and cluster assignments. Inter-coder reliability was assessed using Cohen’s kappa. The observed kappa value for the primary identity category was 0.88, and the categorical visual-cue variables ranged from 0.79 to 0.92, demonstrating strong agreement between coders. Disagreements were resolved through discussion and consensus, and the adjudicated category and visual-cue codes were recorded for subsequent analysis¹². Pre-adjudication labels were preserved.

AI–human agreement and performance estimation
The adjudicated human category was used as the reference for descriptive comparison with one preliminary AI label per painting. Overall exact-match accuracy was calculated as 356/433 (82.2%). Category recall was calculated as the number of exact AI–human matches divided by the human-reference support for each category. Category support, exact matches, mismatches, and corresponding denominators were reported explicitly.

Based on the 433 case-level paired AI and human labels, target-class false positives, per-category precision, recall, and F1 scores were computed. A full confusion matrix was constructed to analyze off-diagonal classification patterns, which subsequently guided the mismatch review, error taxonomy documentation, and adjudication rules.

Reproducibility package and materials
The computational workflow and supporting research materials were organized in a version-controlled repository. The reproducibility package was designed to include scripts for preprocessing, CLIP analysis, SAM localization, UMAP, clustering, agreement analysis, and figure generation, together with the manual coding codebook, configuration files, dependency information, case-level derived outputs, and a machine-readable inventory of image sources and rights information. A persistent identifier was assigned to the archived research materials to facilitate reproducibility and reuse.

Microsoft Excel 2021 was used to organize and manage metadata. Computational analyses used CLIP ViT-B/32 for preliminary vision–language labeling, SAM for visual-element localization, UMAP with cosine distance for dimensionality reduction, and K-means and hierarchical clustering for exploratory pattern analysis. Clustering solutions were evaluated using the silhouette coefficient and Davies–Bouldin index, and inter-coder agreement was assessed using Cohen’s kappa. The software environment, model configurations, computational settings, and random seeds used throughout the workflow were documented to support reproducibility.

Results

Corpus construction and descriptive composition
Figure 1 presents the four-panel overview of the corpus. Figure 1A shows the reported screening process from 612 candidate records to 433 retained records. Figure 1B displays representative examples of media and formats. Figure 1C presents the counts across four date intervals, and Figure 1D summarizes the...

Discussion

Within the supplied aggregate summaries, 433 records were organized into seven human-reference categories and described through five visual-cue families. Elite/beauty was the largest category (112/433), and 356 preliminary labels matched the reported human reference (82.2%). These are corpus observations. They support the study of recurrent pictorial formulas, but not the stronger proposition that the formulas faithfully reproduce Qing society or caused a particular response among foreign audiences¹³

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
CLIP ViT-B/32	OpenAI	OpenAI CLIP, ViT-B/32	Vision-language model used for preliminary image–text similarity analysis and assignment of preliminary AI labels.
Computational workstation	Authors’ computational workstation	CUDA-enabled GPU; exact hardware configuration to be reported	Workstation used for CLIP and SAM inference and computational analyses; GPU, CPU, RAM, operating system, and CUDA specifications should be reported from the original run environment.
Matplotlib	Matplotlib development team	Python package; exact version to be reported	Used for computational visualization and generation of analysis figures and plots.
Microsoft Excel 2021	Microsoft Corporation	Microsoft Excel 2021	Used for metadata organization, screening records, and management of tabular study data.
NumPy	NumPy developers	Python package; exact version to be reported	Used for numerical computation and manipulation of image-feature and analytical arrays.
pandas	pandas development team	Python package; exact version to be reported	Used for structured data handling, metadata processing, and organization of case-level analytical outputs.
Python 3.10	Python Software Foundation	Python 3.10	Programming environment used to implement the computational image-analysis workflow.
PyTorch 2.0.1	PyTorch Foundation	PyTorch 2.0.1	Deep-learning framework used to run CLIP-based image and text encoding and GPU inference.
scikit-learn	scikit-learn developers	Python package; exact version to be reported	Used for K-means clustering, agglomerative clustering, silhouette coefficient, Davies–Bouldin index, and Cohen’s kappa calculations.
Segment Anything Model (SAM)	Meta AI Research	SAM	Segmentation model used to localize female figures, garments, furniture, tools, ritual objects, and architectural regions for visual inspection.
umap-learn	McInnes et al.	UMAP Python implementation	Used for dimensionality reduction and two-dimensional visualization of CLIP image embeddings with cosine distance.