CVNet: confidence voting convolutional neural network for camera spectral sensitivity estimation.

Area of Science:

• Computer Vision
• Digital Imaging
• Machine Learning

Background:

• Spectral sensitivity is crucial for digital camera performance in computer vision.
• Accurate estimation of spectral sensitivity is essential for reliable image analysis.

Purpose of the Study:

• To propose a novel convolutional neural network (CVNet) for reconstructing spectral sensitivity functions.
• To develop a method that utilizes a single image for spectral sensitivity estimation, reducing experimental complexity.

Main Methods:

• A confidence voting convolutional neural network (CVNet) was developed.
• The network models spectral sensitivity as weighted sums of basis functions (Fourier, SVD, Radial).
• Disparate confidence scores were calculated from image segments to learn basis function weights automatically.

Main Results:

• The proposed CVNet achieved high accuracy: 97.92% with Fourier basis functions (FBF), 98.69% with singular value decomposition basis functions (SVDBF), and 99.01% with radial basis functions (RBF).
• The method demonstrated high precision in experimental results.
• The approach was validated through theoretical analysis, dataset creation, and training process demonstration.

Conclusions:

• The proposed CVNet offers a simple, effective, and highly accurate method for spectral sensitivity function estimation.
• This technique could serve as a future alternative to conventional, complex benchtop setups for spectral sensitivity estimation.
• The method's ability to work with a single image and minimal limitations makes it practical for various applications.