In order to improve the controllability and esthetics of product image generation, from the perspective of design, this study proposes a product image generation method based on morphological optimization, esthetic evaluation, and style transfer. Firstly, based on computational esthetics and principles of visual perception, an esthetic comprehensive evaluation model is constructed and used as the fitness function. The genetic algorithm is employed to build a product morphological optimization design system, obtaining product form schemes with higher esthetic quality. Then, an automobile front-end image dataset is constructed, and a generative adversarial network model is trained. Using the aforementioned product form scheme as the content image and selecting automobile front-end images from the market as the target style image, the content features and style features are extracted by the encoder and input into the generator to generate style-transferred images. The discriminator is utilized for judgment, and through iterative optimization, product image schemes that meet the target style are obtained. Experimental results demonstrate that the model generates product images with good effects, showcasing the feasibility of the method and providing robust technical support for intelligent product image design.
To enhance the segmentation accuracy of car front face elements such as headlights and grilles for car front face design, and to improve the superiority and efficiency of solutions in automotive partial modification design, this paper introduces MD-TransUNet, a semantic segmentation network based on the TransUNet model. MD-TransUNet integrates multi-scale attention gates and dynamic-channel graph convolution networks to enhance image restoration across various design drawings. To improve accuracy and detail retention in segmenting automotive front face elements, dynamic-channel graph convolution networks model global channel relationships between contextual sequences, thereby enhancing the Transformer’s channel encoding capabilities. Additionally, a multi-scale attention-based decoder structure is employed to restore feature map dimensions, mitigating the loss of detail in the local feature encoding by the Transformer. Experimental results demonstrate that the MSAG module significantly enhances the model’s ability to capture details, while the DCGCN module improves the segmentation accuracy of the shapes and edges of headlights and grilles. The MD-TransUNet model outperforms existing models on the automotive front face dataset, achieving mF-score, mIoU, and OA metrics of 95.81%, 92.08%, and 98.86%, respectively. Consequently, the MD-TransUNet model increases the precision of automotive front face element segmentation and achieves a more advanced and efficient approach to partial modification design.
1 July 2022 in Computer-Aided Design and Applications
Deep learning has shown strong potential for image‐based aesthetic assessment, but its limited interpretability restricts its use in design‐related applications. This paper proposes XAesViT‐Net, an explainable deep learning framework for car front‐end aesthetic assessment. Inspired by visual aesthetic cognition, the proposed network integrates local feature extraction, global feature modelling and attention‐based feature fusion to improve predictive performance while introducing an interpretability‐oriented architectural design. To evaluate the method, an auto face aesthetics dataset (AFAD) of car front‐view images was constructed using publicly available platform ratings and the task was formulated as a three‐class aesthetic classification. Experimental results show that XAesViT‐Net achieves superior performance against both generic backbones and representative task‐specific aesthetic assessment methods on AFAD, reaching 96.34% accuracy and 0.9972 AUC‐ROC. In addition, Grad‐CAM and SHAP were employed to identify the visual regions and design cues underlying the model predictions, providing post‐hoc interpretability analyses that improve transparency and practical usability. The results demonstrate that the proposed framework is effective for interpretable car front‐end aesthetic assessment. The current validation is limited to car front‐end images and extension to broader product categories remains future work.
As product aesthetics increasingly drive consumer preference, quantitative evaluation remains hindered by subjective evaluation biases and the black-box nature of modern artificial intelligence. This study proposes an advanced machine learning framework incorporating sensitivity-aware morphological features for the aesthetic evaluation of industrial products, with automotive design as a representative case. An aesthetic index system and its quantitative formulations are first developed to capture the morphological characteristics of product form. Subjective weights are determined via grey relational analysis (GRA), while objective weights are calculated using the coefficient of variation method (CVM) integrated with the technique for order preference by similarity to an ideal solution (TOPSIS). A game-theoretic weighting approach is then employed to fuse subjective and objective weights, thereby establishing a multi-scale aesthetic evaluation system. Sensitivity analysis is applied to identify six key indicators, forming a high-quality dataset. To enhance prediction performance, a novel model—improved lung performance-based optimization with backpropagation neural network (ILPOBP)—is proposed, where the optimization process leverages a maximin latin hypercube design (MLHD) to enhance exploration efficiency. The ILPOBP model effectively predicts aesthetic ratings based on limited morphological input data. Experimental results demonstrate that the ILPOBP model outperforms baseline models in terms of accuracy and robustness when handling complex aesthetic information, achieving a significantly lower test set mean absolute relative error (MARE = 4.106%). To further enhance model interpretability, Shapley additive explanations (SHAP) are employed to elucidate the internal decision-making mechanisms, offering reverse design insights for product optimization. The proposed framework offers a novel and effective approach for integrating machine learning into the aesthetic assessment of industrial product design.
A framework for comprehensive aesthetic evaluation of product forms based on images is proposed, emphasizing interpretability and process automation. Starting from visual cognition and visual-organization principles, the method extracts shape-control cues from product images via segmentation and contour processing, and then operationalizes these cues as a set of aesthetic indices. The indices are integrated into a single score using grey relational analysis (GRA), where each weight has an explicit index-level meaning. We evaluate the framework on a car-front case study using large-scale online ratings. By comparing model outputs with large-scale online user ratings, the results indicate that the proposed approach can achieve efficient computation while maintaining a statistically significant positive association with large-scale online ratings. Overall, the framework links principle-oriented graphical descriptions with image-derived measurements, enabling automated assessment with interpretable index contributions and supporting broader applications in product-form design research.