Video flame recognition based on LGATP texture feature and sparse representation

Aiming at the problem of low recognition accuracy caused by inaccurate feature extraction, a video flame detection method based on LGATP (Local Gabor adaptive threshold ternary pattern) texture feature and sparse representation is proposed. Firstly, adaptive local ternary pattern (LTP) is extracted from the multi-directional feature map of Gabor, and more comprehensive detailed texture information is obtained. Secondly, according to the uniform pattern of the local binary pattern, the feature map of the upper and lower pattern is encoded and feature vectors are achieved. Then the feature vectors are cascaded based on information entropy weighted connection to obtain the original LGATP feature. To reduce the calculation complexity, random forest is employed for feature selecting, and the final LGATP texture feature is obtained. Thirdly, a feature dictionary is constructed combined with the features of flame circularity and area change rate. Finally, a weighted kernel sparse representation model is established for flame recognition. The experiment results show that the LGATP texture feature has good feature expression ability and strong robustness, which can improve flame recognition rate effectively.