Riemannian Alternative Matrix Completion for Image-Based Flame Recognition

The flame image has important significance in combustion state recognition and judgment, which can be used effectively for control of energy consumption and exhaust emissions. Due to the harsh industrial environments, flame images are usually corrupted by transmission errors or coding issues, which makes the combustion state analysis very challenging. This paper proposes a novel flame combustion state analysis framework, which provides new insight into two crucial issues: corrupted flame image recovery and combustion state recognition. First, we propose Riemannian alternative optimization (RAO) with fast convergence and the global optimization ability to recover the corrupted flame image. More specifically, RAO constructs a low-rank factorization model and exploits the geometric nature of the flame image to perform the optimization on Riemannian manifolds. Second, we use Fisher discriminant analysis to exploit discriminative features of the recovered flame image and provide well-separated classes of the combustion state for recognition. The experiments show that the proposed framework recovers the corrupted flame image efficiently and achieves satisfying performance of combustion state recognition.