“Bag of visual words” and latent semantic analysis-based burning state recognition for rotary kiln sintering process

摘要

For the sintering process of rotary kiln, the accurate recognition of burning zone state is considered to be the most critical issue. Due to the harsh environment inside the kiln and the limitation of the measuring device, the measurement is still a challenging task. Recently, flame image-based state recognition has received considerate attention. However, the recognition accuracy of previous image segmentation-based methods is hard to guarantee due to the disturbance from smoke and dust. In this study, a new method for burning state recognition without the need of image segmentation is proposed, with the goal of achieving more reliable state recognition. Firstly, scale invariant feature transform (SIFT) operator is employed to extract key feature points of flame image, and then “bag of visual words” is applied to vector quantize the SIFT descriptors, and term frequency-inverse document frequency weight is used to form the indexing table to reduce the dimensionality of feature representation. After obtaining such table, latent semantic analysis (LSA) is used to map the original “images-visual words” space to a latent semantic space to mitigate the problem of synonymy. Previously, very little attention has been paid to the saliency of topics. In our work, a topic selection procedure based on Mahalanobis separability measure is proposed, with the goal of making up the lack of location information to select topics that possess the maximum discriminative power to enhance classification performance. The contribution of our new burning state recognition method is threefold. Firstly, SIFT descriptor is robust to characterize local zones of flame image than the features extracted from image segmentation-based methods. Secondly, “bag of visual words” representation for flame images combined with LSA is feasible to recognize the burning state which has never been used before. Thirdly, our topic selection approach is not only to --generate a more meaningful topic subset, but also to improve classification performance. The proposed new method is validated through extensive experimental studies.