Enhanced monocular depth estimation using novel scale-invariant Error Structure Similarity Index measure optimization in Convolutional Neural network architecture

Monocular Depth Estimation (MDE) is crucial for applications like autonomous driving, medical imaging, and 3D modeling. This paper presents a novel Convolutional Neural Network (CNN) architecture that balances performance and computational cost in MDE tasks. Key components include bottleneck mechanisms, Modified Convolutional Block Attention Module (MCBAM), Atrous Spatial Pyramid Pooling (ASPP), and Pyramid Scene Parsing (PSP). Leveraging pre-trained backbones and attention mechanisms, our model significantly improves depth estimation accuracy and reduces computational complexity. Validated using the NYU Depth Dataset V2, our model outperforms existing benchmarks in Absolute Relative Error (Abs Rel), Square Relative Error (Sq Rel), Root Mean Square Error (RMSE), and Thresholding metrics. A novel loss function incorporating Structure Similarity Index Measure (SSIM) and Scale-Invariant Error (SIE) enhances training and evaluation. Our study advances MDE techniques, offering a practical solution with wide-ranging applications. Future research will explore attention mechanisms, fusion approaches, and real-time optimization for greater versatility.