Improving Video Temporal Consistency via Broad Learning System.

Area of Science:

• Computer Vision
• Video Processing
• Machine Learning

Background:

• Image-based processing on videos can disrupt temporal consistency between frames, leading to flickers.
• Existing methods for video temporal consistency often rely on optical flow, which can be inaccurate and limit practical application.

Purpose of the Study:

• To propose a novel approach, the Temporally Broad Learning System (TBLS), for enforcing temporal consistency in videos.
• To address the limitations of traditional methods in reconstructing temporally consistent frames and maintaining video fidelity.

Main Methods:

• The TBLS is designed as a flat network processing original and temporally inconsistent frames.
• It refines features using enhancement nodes with random weights and connects them to an output layer with a target weight vector.
• An incremental learning algorithm is proposed to enhance learning accuracy through broad expansion.

Main Results:

• The TBLS successfully minimizes temporal information loss and video fidelity loss.
• Experimental results demonstrate the superiority of the proposed TBLS in removing temporal inconsistency.
• The system outputs temporally consistent videos after processing.

Conclusions:

• The Temporally Broad Learning System (TBLS) effectively restores temporal consistency in videos processed with image-based methods.
• The proposed method overcomes the accuracy limitations of optical flow in traditional approaches.
• TBLS offers a robust solution for enhancing video quality and temporal coherence.