Quality assessment of real estate photographs: A deep learning-based approach

In the digital era, where first impressions are often formed through online visuals, the quality of real estate photographs plays a crucial role in influencing prospective buyers' decisions and shaping their perception of properties. Visual presentation in real estate listings is no longer a supplementary feature but a vital determinant of user engagement and transaction success. Numerous studies have revealed that high-quality images can significantly increase the perceived value of a property and drive user interactions, while low-quality visuals may reduce trust and interest. This study proposes an automated system for real estate image quality assessment using deep learning methodologies. While previous research in image quality evaluation has focused primarily on aesthetic judgments, generic image enhancement, or low-level perceptual metrics, the present study addresses a specific and practical need: assessing and filtering real estate photographs based on their visual adequacy for online listing platforms. The work is rooted in the recognition that the manual moderation of thousands of user-uploaded images is labor-intensive, subjective, and unsustainable for large-scale real estate platforms. Hence, automation in quality control using artificial intelligence (AI) provides a scalable and consistent alternative. To this end, three convolutional neural network (CNN) models—MobileNetV2, ResNet18, and EfficientNetB1—were selected for comparative analysis based on their architectural diversity, computational complexity, and prior success in vision-based tasks. These models represent different design philosophies: MobileNetV2 emphasizes lightweight design and efficiency, ResNet18 utilizes residual connections for deeper feature extraction, and EfficientNetB1 introduces compound scaling to balance model accuracy and complexity. The dataset for this study comprises 6,330 images collected from a major real estate platform in Turkey, representing real user uploads. Each image was manually labeled as either "Good" or "Bad" based on clear criteria such as sharpness, lighting, absence of watermarks, and visual clarity. The "Good" class includes professionally appropriate photographs, while the "Bad" class includes blurry, poorly lit, overexposed, or logo-marked images. The dataset was split into training (4,446), validation (628), and testing (1,256) subsets. This manual labeling process was meticulously carried out by the author and real estate platform employees to ensure objectivity and reliability in model training. Preprocessing steps included image resizing to 224x224 pixels and pixel normalization to the [0,1] range. Additionally, data augmentation techniques—such as random flipping, cropping, and brightness adjustments—were applied during training to enhance the model's generalization capability. However, experimental results showed marginal performance improvements with augmentation; hence, models were primarily trained without these enhancements for optimal efficiency.A transfer learning strategy was adopted to leverage the pretrained knowledge embedded in ImageNet-based architectures. The convolutional base of each model was frozen, and only the classification head was fine-tuned on the custom real estate dataset. This approach significantly reduced training time and prevented overfitting, making it especially suitable for domain-specific tasks with moderate-sized datasets. Performance evaluation was conducted using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices. Among the three models, MobileNetV2 consistently outperformed its counterparts across multiple metrics. It achieved an accuracy of 89.17%, an AUC of 0.9435, a precision of 0.9424, a recall of 0.8344, and an F1-score of 0.8851. These results underscore MobileNetV2's capability to balance computational efficiency and classification performance, making it the most suitable model for real-time applications on listing platforms. The confusion matrix analysis revealed that MobileNetV2 had the lowest false positive and false negative rates, especially in identifying low-quality images. On the other hand, ResNet18 demonstrated robust recall and moderately high precision but suffered from slightly lower accuracy and higher training time. EfficientNetB1, while maintaining good precision and F1-scores, exhibited lower recall in detecting poor-quality images. This suggests that while it is effective in identifying good images, it is less sensitive to problematic visual attributes. The findings from this research provide concrete evidence that lightweight CNN architectures such as MobileNetV2 can be effectively integrated into the upload pipelines of real estate platforms. By embedding such models at the point of content submission, platforms can guide users in real-time to upload only acceptable quality visuals. This not only improves the aesthetic and informational quality of listings but also reduces the need for manual moderation. Moreover, maintaining consistent visual standards contributes to brand credibility and user trust, ultimately enhancing the platform's competitiveness in the market. In practical deployment scenarios, a simplified variant of the best-performing model was developed to classify single image inputs as either "High Quality" or "Low Quality." This enables instant feedback to users, streamlining the content validation process and preventing the publication of substandard visuals. Beyond the immediate benefits to listing platforms, the implementation of such an automated system also opens avenues for broader applications within the real estate industry. For example, quality assessment models can be integrated into agent-facing tools to help realtors curate their portfolios more effectively or to suggest optimal photo-taking practices. Additionally, these models can contribute to market analytics by enabling large-scale analysis of photo quality trends over time and across regions, offering insights into how visual presentation correlates with sales velocity and pricing strategies. Furthermore, incorporating explainable AI techniques into the quality assessment pipeline represents an exciting direction for future research. By providing interpretable feedback—such as highlighting blurry areas or poorly lit regions—users can gain actionable insights into how to improve their photographs. This transparency fosters user trust and facilitates educational opportunities, encouraging users to adhere to best practices while also enhancing the model's acceptance and effectiveness within real-world workflows.Future work could explore the inclusion of additional quality dimensions, such as composition, color harmony, or presence of human subjects, to refine classification granularity. Furthermore, expanding the dataset to include international listings and multilingual interfaces could enable cross-cultural model training, enhancing generalizability. Lastly, an end-to-end quality control framework combining deep learning with explainable AI techniques could help platform moderators understand model decisions and ensure transparency in automated evaluations. In conclusion, this thesis demonstrates that real estate photo quality assessment can be reliably and efficiently automated using deep learning, specifically with MobileNetV2. The proposed approach offers a foundational step toward integrating AI-based visual quality control mechanisms into digital real estate ecosystems, promising enhanced user experience, operational efficiency, and market competitiveness.