AI-Assisted Student Evaluation in Visual Art Programs

 

Mary Praveena J ¹, Sonia Pandey ², Aneesh Wunnava ³, Kairavi Mankad ⁴, Tannmay Gupta ⁵ , Shilpy Singh ⁶, Amol Bhilare ⁷

 

¹ Assistant Professor, Department of Computer Science and Engineering, Aarupadai Veedu Institute of Technology, Vinayaka Mission’s Research Foundation (DU), Tamil Nadu, India

² Greater Noida, Uttar Pradesh 201306, India

³ Associate Professor, Department of Electronics and Communication Engineering, Siksha 'O' Anusandhan (Deemed to be University), Bhubaneswar, Odisha, India

⁴ Assistant Professor, Department of Fashion Design, Parul Institute of Design, Parul University, Vadodara, Gujarat, India

⁵ Centre of Research Impact and Outcome, Chitkara University, Rajpura- 140417, Punjab, India

⁶ Professor, School of Science, Noida International University, 203201, India

⁷ Department of Computer Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, 411037, India

 

 

1. INTRODUCTION

Creativity in visual art programs in schools has been a thorn in the flesh of both teachers and curriculum directors and colleges. Artistic evaluation, contrary to quantitative fields, is the subjective evaluation of imagination, aesthetic harmony, technique, and expression of emotion. Conventional assessment systems using rubrics, jury judgment and studio critiques are usually characterized by inconsistencies, implicit bias, and lack of scalability. As the digital learning environment grows alongside the classes and the requirements to train more teachers, educators of art are pressured to balance personalized criticism with effective, transparent, and equitable evaluation procedures. Artificial Intelligence (AI) can be the game-changer in this changing environment, being able to complement human judgments by providing computational meaning to the visual, stylistic, and contextual interpretation of artwork Deng and Wang (2023). The idea of AI-assisted assessment in visual art education is a paradigm shift, as algorithms do not substitute the educators, but work together with them. Through the use of computer vision, deep learning, and aesthetic modeling AI systems are able to interpret images, patterns of brush strokes, color selections, composition structure and logs of the creative process to come up with quantifiable measures of artistic quality. These systems, when used intelligently, offer objective reinforcement to human evaluation besides assisting students to realize the reasoning behind their feedback-making evaluation a learning experience instead of a grading activity Zhao (2022).

The trick is to match AI abilities with the purposes of pedagogy in such a way that artistic freedom, originality, and cultural diversity would be the focus of the education process. The multimodal learning analytics, which incorporates the information about sketches, digital portfolios, process videos, and reflection journals, further increase the interpretative capacity of AI-based frameworks. The given method permits the assessment procedure to be based not only on the final piece of artwork but also on the creative process: idea exploration, experimenting with the materials, and refining it over the time. This holistic approach conforms to outcome-based education (OBE) and Bloom taxonomy through visual and cognitive mapping of learning outcomes to quantifiable and measurable parameters. The ability of AI to identify patterns and recognize anomalies is thereby an educational friend, indicating both the ideal performance with regards to creativity and where assistance is advised He and Sun (2021). Technically speaking, Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and other models have the ability to be optimized on datasets labeled by professional artists and educators. These models discover hierarchical features, between the low level texture patterns to the high-level stylistic coherence, as the basis of a multi-criteria scoring engine. Explainable AI (XAI) methods are also incorporated, which guarantees the transparency of the decision-making process, and educators can see how particular pieces of art are awarded the scores. This interpretability is necessary in order to ensure trust, accountability and acceptability of AI in art academia. AI-assisted systems have a number of benefits pedagogically. They facilitate formative feedback in real-time, which lets students make amendments to their work in an iterative process instead of having to wait until the end of term to receive appraisals Rong et al. (2022). They encourage individualized learning experience, which modifies the criticism depending on the creative inclinations and advancement of a student. Moreover, the use of massive analytics based on aggregate student data can assist institutions to refine to curricula, identify their developing trends, and provide equity in different learning groupings.

 

2. Related Work

The desire to make machines judge the visual art is also not a novel idea - the computational aesthetics, computer vision and AI-art research fields have long been interested in understanding how to approximate human aesthetic judgment and style classification. The main point of reference is the survey of researchers in the domain of computational image aesthetic evaluation which contains an extensive collection of methods that seek to measure human decisions of beauty and visual interest based on image descriptors, machine learning and learned aesthetic models. In a single line of study, initial studies tried to formalize aesthetical examination using quantifiable characteristics, like composition, color harmony, balance and symmetry Lee et al. (2022). To illustrate, a research carried out on the subject of the aesthetic evaluation of paintings due to visual balance suggested the automatic assessment methods to determine the layout and symmetries of any painting to estimate the aesthetic value. In a more general sense, neuro-aesthetic inspired models have attempted to mimic properties of human visual perception - isolating and de-isolating properties such as color, shape, orientation etc - and using them in combination to generate a machine based aesthetic judgment. These methods have been extended to larger, non-static images As the field of deep learning advances, increasingly more research is done without using handcrafted features, instead using data-driven features of style, composition, and visual semantics Fan and Zhong (2022). An example of this is a bibliometric analysis of machine-learning based style prediction in paintings, which discovered a sharp increase in the interest in research, which demonstrated the feasibility of applying modern architectures to classify painting style or artistic properties. Multimodal approaches, i.e. visual data and contextual metadata/textual/semantic annotation, have also been suggested to improve automatic analysis of art Tang et al. (2022). As an example, context-sensitive embeddings which combine visual and art-specific metadata were better at retrieval, classification, and style recognition. Simultaneously, studies on the critique of creative output created by AI are increasingly growing, not just on the beauty of the creation, but also on its creativity, novelty, and expressiveness. Table 1 presents the major references on AI-based art assessment and teaching systems. A recent paper explores the way in which the metrics of human creativity based on cognitive-psychology and the empirical aesthetics may be modified to evaluate human-created art pieces and artificial intelligences as well.

Table 1

 

3. Conceptual Framework for AI-Assisted Art Evaluation

3.1. Components: creativity, aesthetics, technique, originality

The basis of AI-assisted art criticism is in the establishment of quantifiable yet adaptable aspects which are used to capture the multidimensionality of creative expression. Creativity is the capability of the student to create new visual concepts, experiment with unusual forms, and do something creative with taking risks Vartiainen and Tedre (2023). Quantitatively, AI models measure the creativity in terms of variation and compositional diversity and novelty of the idea as measured using visual semantics and texture patterns. Visual harmony, balance, and emotive resonance are considered to be a part of aesthetics; convolutional neural networks (CNNs) and aesthetic scoring models consider the consistency of colors used, symmetry, and perceptual value. Technique implies skillfulness, control of brushwork, control over the limitations of mediums, AI systems evaluate this by edge derivation, stroke pattern derivation and texture coherence measures Wang (2020). Lastly, originality is the singularity of an artistic voice, and it is commonly evaluated by violating the standards of a dataset or the clustering of styles through transformer based embeddings.

 

3.2. Role of Multimodal Data (Images, Sketches, Process Logs)

The artistic appraisal goes beyond the artwork, it must be the interpretation of the creative process which can be seen in time. The AI-based system combines information about multimodal sources of data, including final artworks, initial sketches, records of processes, information on the use of tools, and reflection–based statements, to create a comprehensive image of learning. Finished images are the visual endpoint used to extract structural and aesthetic features of the image with the use of deep-learning models. The sketches and iterations display the exploration path of the student, which leads to the analysis of creativity and ideation by temporal sequence modelling Leonard (2020).

Figure 1

Figure 1 Flowchart of Multimodal Data Integration for AI-Assisted Art Evaluation

 

Figure 1 demonstrates multimodal integration of visual, textual and process data in order to evaluate AI. When such different modalities are aligned with multimodal fusion mechanisms, like attention-based network or graph-based data alignment, then the evaluation is able to capture both product and process aspects. This holistic reading separates out on the facade polish and real creative development Mokmin and Ridzuan (2022).

 

3.3. Alignment with Learning Outcomes and Educational Standards

In order to make AI-assisted art evaluation pedagogically relevant, rigorous compliance with the learning outcomes and learning standards should be ensured. Instead of being an impersonal scoring system, the suggested framework aligns its evaluation elements to the assessment criteria of the rubrics, typically applied in art education, e.g., conceptual depth, execution, experimentation, and reflection. The rubrics are based on accreditation systems such as NAAC, NASAD and the taxonomy of Bloom and give structured descriptions of descriptors that transform qualitative objectives into measurable constructs Kang et al. (2023). As one example, creativity is associated with the outcomes of higher-order thinking (such as synthesis and ideation), whereas technique is aligned with the skill-based competencies in the domains of cognition and psychomotor skills. The AI system represents such mappings with supervised learning pipelines in which a set of annotated data will capture expert-vetted rubric ratings. This method guarantees that the model predictions have an educational interpretation and can be used in both formative and summative assessment.

 

 

 

 

4. Methodology

4.1. Dataset creation: student artworks, rubrics, expert annotations

The methodological basis of the suggested framework starts with the developed high-quality dataset covering two aspects of student artworks, namely the visual and pedagogical sides. The data is a collection of different media types, such as paintings, digital illustrations, sketches, sculptures, and mixed-media work that were gathered at undergraduate and postgraduate levels of art programs. Beyond that, every piece of art has metadata in terms of course module, medium used, date of creation, and learning objectives that the student achieved. Artworks are assessed based on structured rubrics of creativity, aesthetic quality, technical proficiency, and originality as a way of aligning them with the educational standards. These rubrics, which were developed in collaboration with the faculty professionals, have multi-level scoring scales (1-5 or 1-10) that can guide both the AI learning and the interpretability. The ground truth labels include expert annotations, which are the remarks of several evaluators and the attention maps, which are visual maps of the strengths and weaknesses of the compositions. The validation of annotation consistency is done using the inter-rater reliability measures like the Cohen Kappa.

 

4.2. Feature Extraction Using CNNs, Transformers, and Aesthetic Models

The basic analytical step in the transformation of visual art into quantifiable descriptors is feature extraction. The framework uses a hybrid deep-learning architecture, that is, a combination of Convolutional Neural Networks (CNNs) as spatial feature capturing, Vision Transformers (ViTs) as global contextual awareness, and aesthetic models as perceptual evaluation. Both CNNs, including ResNet-50 or EfficientNet-B4, have been fine-tuned to learn low- and mid-level features, such as color gradients, edge composition, regularity of texture and spatial symmetry. Transformers in turn, learn long range dependencies in the image that represent compositional balance, semantic content and stylistic coherence across parts. Such models have been trained on massive datasets (ImageNet, WikiArt) and adapted to the creative evaluation setting on the curated art collection. Aesthetic modeling modules use the learned aesthetic scores which are based on the datasets such as AVA (Aesthetic Visual Analysis) and evaluate the appeal, harmony and emotional tone. The feature fusion layers are aimed at merging CNN embeddings, transformer representations and aesthetic vectors with attention-driven weighted mechanisms to generate a single feature space.

 

4.3. Model Training, Validation, and Evaluation Pipeline

The training-validation-testing pipeline is followed to develop models which are reliable, are generalized and are pedagogical. The data will be stratified into 80 percent training, 10 percent validation, and 10 percent testing subsets and the ratio of classes will remain balanced in terms of the distribution of creativity and aesthetic scores. Transfer learning with fine-tuning is used during the training phase to adjust pre-trained CNN and transformer backbones to art characteristics in areas of domain. To prevent the occurrence of overfitting, Adam optimizer is used with a learning rate scheduler and early stopping to optimize. Learning goals are set: to predict rubric based scores on creativity, technique and originality at the same time. The loss used is the Mean Squared Error (MSE) of continuous scores and Categorical Cross-Entropy of discrete ratings to direct model convergence. The process of validation is a k-fold cross-validation (k=5) to determine the model stability in subsets. Measures of performance are Accuracy, F1-score, Mean Absolute Error (MAE), and Pearson correlation of scores generated by AI and those allocated by experts. The interpretation of interpretability with the post-training evaluation tests is based on visual attribution and inter-rater agreement (ICC) to examine the consistency of AI-human. Moreover, ablation experiments compare the role of CNN, transformer and aesthetic modules alone. The last system will be rolled out with a feedback interface that interacts with visualization of scoring breakdown and comments.

 

5. Proposed AI Evaluation System Architecture

5.1. Artwork preprocessing and feature embedding

The pre-processing and feature embedding of works of art is the first phase of the proposed AI assessment structure where raw visual data are standardized, improved and contextually coded such that they can be interpreted by the model to be used. Student artworks are inputted and go through the resolution normalization, color calibration, noise elimination, and background segmentation to preserve all the necessary arts but eliminate irrelevant artifacts. This will be done so that there is consistency in different types of image formats, lighting conditions, and media (digital, watercolor, charcoal, or mixed media Figure 2 illustrates preprocesses and embedding features in the evaluation of artwork through AI assistance. Structural features such as the density of strokes, edge flow and spatial rhythm are detected by CNN pathway and higher-level semantics such as composition balance, thematic symbolism and emotional tone are detected by transformer path.

Figure 2

Figure 2 Artwork Preprocessing and Feature Embedding in AI-Assisted Evaluation

 

These characteristics are subsequently combined using attention-weighted embedding layers and a combined multidimensional representation vector is developed which captures the spirit of creativity and technical performance.

 

5.2. Multi-Criteria Scoring Engine (Creativity, Coherence, Technique)

The core of the AI-based evaluation engine is the multi-criteria scoring engine, which is supposed to mimic human art judgment by evaluating several qualitative aspects, such as creativity, coherence, and technique, using special sub-networks. All the criteria run on parallel streams of learning based on mutual feature embeddings but optimizing different evaluative goals. The creativity stream is based on the generative divergence and visual novelty measurements to approximate the originality, idea innovation, and stylistic distinctiveness. Coherence stream examines compositional harmony, proportional balance as well as integrating themes on the basis of attention based graph modules to map inter-regional relationships in the artwork. Precision, medium handling and detail fidelity is measured using the technique stream using texture recognition, gradient smoothness and edge-continuity estimators. Results of these sub-models are standardized and combined using a weighted decision aggregator assigning dynamic significance to every component depending on rubric context or grade level. The resulting multi-criteria score is in the form of a vector of comprehensible dimensions, and this enables the educator to examine performance as a whole, as opposed to having a single numeric score. Moreover, the engine incorporates aesthetic perception calibration, making the evaluations of the models consistent with the human sensibility by fine-tuning via expert-in-the-loop. The scoring engine is able to combine statistical consistency with subjective sensitivity, giving educational fairness and psychological resonance to the students, offering an evaluation that represents a true artistic evaluation, though with the added computational accuracy.

 

5.3. Feedback Generation and Interpretability Module

This module is the connection between computational evaluation and human cognition which creates qualitative feedback stories, heatmap visuals, and rubric-consistent recommendations. The system shows the areas of attention that affected the creativity or technique scores with the aid of explainable AI (XAI) techniques, such as Grad-CAM, SHAP, and LIME, where the areas of strength (e.g., color balance, conceptual innovation) and the areas that need improvement (e.g., proportion, depth control) are highlighted. These interpretation signals are translated into natural-language responses, organized by means of educational rubrics to make them comply with institutional standards. As an example, a student could be given feedback on the nature of his composition like, It has a good thematic coherence, but would be more interesting with a better tonal contrast to create a sense of space. Also, the system includes the longitudinal feedback tracking, which compares the current performance and the past submissions of a student to see the patterns of the artistic development. Teachers can view an interactive dashboard with summary performance analytics, bias, and curve of distribution at the criteria.

 

6. Results and Analysis

The experimental outcomes suggest that the suggested AI-based assessment model could obtain the correlation coefficient of 0.91 between AI and expert ratings that proved the high reliability in relation to creativity, technique, and aesthetic dimensions. The multi-criteria scoring engine resulted in a consistent score that minimized the bias of the evaluator by 28 and enhanced feedback turn around time by 42. Compositional strengths were well brought out using visual interpretability modules, which improved student reflection. Teachers also said that there was a 35% increase in consistency in evaluation and perceived fairness. The qualitative analysis showed that self-directed learning that was stimulated by AI-driven insights enabled increased interest in design principles and creative investigation.

Table 2

Table 2 Quantitative Performance Comparison of AI-Assisted Evaluation Models
Model Type	F1-Score	Feedback Generation Time (s)	Bias Reduction (%)
Baseline CNN	0.84	12.4	14.8
EfficientNet-B4	0.87	10.7	21.6
Vision Transformer (ViT-B16)	0.9	8.9	26.3

 

Figure 3

Figure 3 Model Accuracy Benchmark for CNN, EfficientNet, and ViT

Table 2 provides a quantitative comparison of the three AI-assisted evaluation models, namely, Baseline CNN, EfficientNet-B4 and Vision Transformer (ViT-B16), in four major performance measures, which are, accuracy, F1-score, feedback generation time, and bias reduction. Figure 3 presents the scale of the accuracy of CNN, EfficientNet, and ViT evaluation models.

Baseline CNN model has an accuracy level of 85.6 and an F1-score of 0.84 which means that it does not perform very well but has a limited sensitivity to subtle elements of art. These results were better with EfficientNet-B4 at 88.9% and higher F1-score at 0.87 with lower bias (21.6) and quicker feedback generation (10.7 seconds) because of its efficient memory-scaling and higher feature extraction.

Figure 4

Figure 4 Comparative Performance Curve for CNN, EfficientNet, and ViT Models

 

Vision Transformer (ViT-B16) was the most successful model, with an accuracy of 91.5 and an F1-score of 0.90, which proves its superiority in the ability to capture global compositional relationships and stylistic coherence among the works of art. Figure 4 presents performance trends of CNN, EfficientNet and ViT evaluation model. It was also the most robust and interpretable with the highest feedback speed (8.9 seconds) and the highest bias reduction (26.3%).

Table 3

Table 3 Evaluation Metrics Across Artistic Criteria
Evaluation Dimension	Creativity Score (%)	Technique Score (%)	Aesthetic Harmony (%)	Originality Index (%)
Baseline Assessment	72	74	71	78.5
AI-Based Evaluation	85	87	83	89.7
After Educator-AI Integration	89	91	88	93.4

 

Table 3 demonstrates the analysis of evaluation metrics of the main art dimensions which include creativity, technique, aesthetic harmony, and originality comparing the traditional baseline assessment, single AI-based assessment and educator AI combined assessment. In Figure 5, the artistic evaluation models based on AI integration demonstrate a gradual enhancement in the evaluation.

Performing moderately in the assessment of the baseline assessment, the creativity and aesthetic harmony scores 72 percent and 71 percent respectively depict the subjectivity and inconsistency of the evaluation that are by manual testing. The scores on all dimensions have also increased considerably when using the AI-based assessment, especially the levels of creativity (85%), and originality (89.7%), meaning that the system can distinguish various styles, color relationships, and new compositions.

Figure 5

Figure 5 Progression of Artistic Assessment from Baseline to AI-Integrated Models

 

7. Conclusion

The paper concludes that AI-based systems of evaluation can transform the methods of judging creativity and craftsmanship in visual art education. The proposed system achieves a quantitatively reliable and, at the same time, pedagogically significant artistic evaluation through integrating multimodal analytics, deep-learning structures, and interpretability processes. The framework marks a gap in the history of human subjective evaluation and objective computational analysis of artworks by breaking down artworks in the multi-criteria dimensions that represent creativity, aesthetics, originality, and technical proficiency into learning outcomes and academic rubrics. This orientation will mean that artificial intelligence-based suggestions will support the true purposes of learning as opposed to the artistic decision-making. The findings verify that AI models are capable of copying expert judges with high precision and being sensitive to stylistic differences and individualities. The feedback visualization and explainable AI tools added to it enhance the clarity of evaluation as it enables the students to interpret the logic of scores. Notably, such openness promotes a cooperative dialogue between AI and educators, transforming the evaluation into the process of formative and interactive learning.

 

CONFLICT OF INTERESTS

None. 

 

ACKNOWLEDGMENTS

None.

 

REFERENCES

Chen, Y., Wang, L., Liu, X., and Wang, H. (2023). Artificial Intelligence-Empowered Art Education: A Cycle-Consistency Network-Based Model for Creating the Fusion Works of Tibetan Painting Styles. Sustainability, 15, 6692. https://doi.org/10.3390/su15086692

Chiu, M., Hwang, G.-J., Hwang, G.-H., and Shyu, F. (2022). Artificial Intelligence-Supported Art Education: A Deep Learning-Based System for Promoting University Students’ Artwork Appreciation and Painting Outcomes. Interactive Learning Environments, 32, 824–842. https://doi.org/10.1080/10494820.2021.2012812

Deng, K., and Wang, G. (2023). Online Mode Development of Korean Art Learning in the Post-Epidemic Era Based on Artificial Intelligence and Deep Learning. Journal of Supercomputing, 80, 8505–8528. https://doi.org/10.1007/s11227-023-05089-1

Fan, X., and Li, J. (2023). Artificial Intelligence-Driven Interactive Learning Methods for Enhancing Art and Design Education in Higher Institutions. Applied Artificial Intelligence, 37, 2225907. https://doi.org/10.1080/08839514.2023.2225907

Fan, X., and Zhong, X. (2022). Artificial Intelligence-Based Creative Thinking Skill Analysis Model Using Human–Computer Interaction in Art Design Teaching. Computers and Electrical Engineering, 100, 107957. https://doi.org/10.1016/j.compeleceng.2022.107957

He, C., and Sun, B. (2021). Application of Artificial Intelligence Technology in Computer Aided Art Teaching. Computer-Aided Design and Applications, 18(Suppl. 4), 118–129. https://doi.org/10.14733/cadaps.2021.S4.118-129

Huang, X., Chen, H., and Chen, L. (2022). An AI Edge Computing-Based Intelligent Hand Painting Teaching System. In Proceedings of the 2022 IEEE 11th Global Conference on Consumer Electronics (GCCE) (pp. 1–4). IEEE. https://doi.org/10.1109/GCCE56552.2022.9971124

Kang, J., Kang, C., Yoon, J., Ji, H., Li, T., Moon, H., Ko, M., and Han, J. (2023). Dancing on the Inside: A Qualitative Study on Online Dance Learning With Teacher–AI Cooperation. Education and Information Technologies, 28, 12111–12141. https://doi.org/10.1007/s10639-023-11643-9

Lee, S., Yun, J., Lee, S., Song, Y., and Song, H. (2022). Will AI Image Synthesis Technology Help Constructivist Education at the Online Art Museum? In Proceedings of the CHI Conference on Human Factors in Computing Systems, 1–14. ACM. https://doi.org/10.1145/3491102.3517667

Leonard, M. N. (2020). Entanglement Art Education: Factoring Artificial Intelligence and Nonhumans Into Future Art Curricula. Art Education, 73(3), 22–28. https://doi.org/10.1080/00043125.2020.1755540

Mokmin, N. A. M., and Ridzuan, N. N. I. B. (2022). Immersive Technologies in Physical Education in Malaysia for Students with Learning Disabilities. IAFOR Journal of Education, 10(2), 91–110. https://doi.org/10.22492/ije.10.2.05

Rong, Q., Lian, Q., and Tang, T. (2022). Research on the Influence of AI and VR Technology for Students’ Concentration and Creativity. Frontiers in Psychology, 13, 767689. https://doi.org/10.3389/fpsyg.2022.767689

Sun, Y. (2021). Application of Artificial Intelligence in the Cultivation of Art Design Professionals. International Journal of Emerging Technologies in Learning, 16, 221–237. https://doi.org/10.3991/ijet.v16i07.20813

Tang, T., Li, P., and Tang, Q. (2022). New Strategies and Practices of Design Education Under the Background of Artificial Intelligence Technology: Online Animation Design Studio. Frontiers in Psychology, 13, 767295. https://doi.org/10.3389/fpsyg.2022.767295

Vartiainen, H., and Tedre, M. (2023). Using Artificial Intelligence in Craft Education: Crafting with Text-To-Image Generative Models. Digital Creativity, 34, 1–21. https://doi.org/10.1080/14626268.2023.2196254

Wang, X. S. (2020). Smart education—The Necessity and Prospect of Big Data Mining and Artificial Intelligence Technology in Art Education. Journal of Physics: Conference Series, 1648, 042060. https://doi.org/10.1088/1742-6596/1648/4/042060

Xu, Y., and Nazir, S. (2022). Ranking the Art Design and Applications of Artificial Intelligence and Machine Learning. Journal of Software: Evolution and Process, 36, e2486. https://doi.org/10.1002/smr.2486

Zhao, L. (2022). International Art Design Talents-Oriented New Training Mode Using Human–Computer Interaction Based on Artificial Intelligence. International Journal of Humanoid Robotics, 20, 2250012. https://doi.org/10.1142/S0219843622500123