Jitter Elimination in Shape Recovery by using Adaptive Neural Network Filter

Area of Science:

• Computer Vision
• Optical Engineering
• Image Processing

Background:

• Three-dimensional (3D) cameras are often costly due to complex sensors and optics.
• Shape from Focus (SFF) offers a cost-effective passive optical approach for 3D imaging.
• Mechanical vibrations in SFF systems introduce jitter noise, degrading 3D shape accuracy.

Purpose of the Study:

• To develop an accurate depth estimation method for SFF 3D imaging systems.
• To mitigate the detrimental effects of jitter noise on 3D shape recovery.
• To improve the accuracy of low-cost 3D cameras utilizing SFF techniques.

Main Methods:

• An adaptive neural network (ANN) filter was designed as an optimal estimator.
• The ANN filter preprocesses image sequences to remove jitter noise effects.
• Jitter noise was modeled using both Gaussian and non-Gaussian distributions, with focus curves modeled by quadratic functions.

Main Results:

• The proposed ANN filter effectively removes jitter noise from SFF image sequences.
• Depth estimation accuracy was significantly improved compared to traditional methods.
• Experimental results with synthetic and real objects demonstrated comparable accuracy to existing systems.

Conclusions:

• The adaptive neural network filter provides an efficient and accurate solution for jitter noise in SFF 3D imaging.
• This method enhances the reliability of depth estimation in low-cost 3D camera systems.
• The approach offers a viable alternative for achieving accurate 3D shape information despite mechanical vibrations.