Research Information System
6th Mediterranean Conference on Pattern Recognition and Artificial Intelligence, MedPRAI 2024, İstanbul, Turkey, 18 - 19 October 2024, vol.1393 LNNS, pp.1207-1222, (Full Text)
This study aims to accurately forecast depth maps and estimate volume from 2D images. Medical imaging, robotics, and computer vision all have the issue of prediction using 2D data. Traditional methods have struggled to generalize across many scenarios and datasets. To overcome these limitations, a novel deep learning-feature extraction method has been devised. A U-Net model predicted depth maps more accurately by capturing complicated spatial hierarchies with its powerful convolutional network architecture. HOG (Histogram of Oriented Gradients) and Oriented FAST, Rotated BRIEF (ORB) feature extraction enhanced the model’s object volume estimation. A custom Blender-generated dataset and the NYU Depth V2 dataset were used to validate the recommended algorithms. In testing using Random Forest, Support Vector Regression (SVR), and Gradient Boosting Machine (GBM), the recommended technique outperformed them. Improvements in the R-squared (R2) and Mean Squared Error (MSE) metrics are observed. The results demonstrate the effectiveness of deep learning with traditional feature extraction and regression models, paving the way for more accurate volume estimation from 2D images. This study discusses a strong framework that enhances depth and volume estimate accuracy, and the framework is scalable and domain-adaptable. The findings indicate that the proposed approach may be used for other applications that need precise 3D reconstructions from 2D data, which is significant for future study.