Abstract: Forest fire image recognition/detection plays a vital role in forest fire monitoring system. Due to its own characteristics and difficulties of forest fire image, the existing studies mainly focus on the vector-pattern-oriented fire image, where each vector-pattern sample corresponds to an image pixel one by one. Since the number of vector-pattern samples is strongly determined by the resolution of the given image, it is time-consuming for training classifier to deal with numerous vector-pattern samples, especially for higher-quality images. How to label samples is another big challenge in the task of image target recognition. However, at present, this labeling work is done manually or semi-manually (for instance, the method of image preprocessing). It is clear that the accuracy of labels directly affects subsequent steps including classifier training and object recognition. Furthermore, owing to the rearrangement of adjacency relationship between pixels, vector-pattern samples, which are generated from image pixel-by-pixel vectorization, unavoidably lost the original image structural information. In this paper, we proposed a matrix-pattern semi-supervised algorithm for forest fire image recognition, named Semi-MHKS (semi-supervised matrix-pattern Ho-Koshyap algorithm with squared approximation). Its advantages lie in 4 aspects: ①Instead of vector-pattern, it adopts sub-matrix-pattern samples to train classifier. In doing so, it is more likely to meet real-time requirements because of smaller size of training set. ②It is easier to label the training samples in the manner of sub-matrix-pattern than that of vector pattern. Moreover, it is also effective for decreasing the error rate in manual-labeling. ③Adopting so-called bi-linear discriminant function, we design a semi-supervised learning algorithm (Semi-MHKS) for forest fire images, which only needs several labeled samples. It is also suitable for classifying the a batch of unknown matrix-pattern samples. ④The algorithm leads to a strictly convex optimization problem, which can be solved by quadratic programming and gradient descend method. It is mathematically proved that Semi-MHKS is convergent in the stage of alternating iteration, with fixed left or right weight vectors of the bi-linear function. Compared to state-of-the-art methods, including vector-pattern support vector machine (SVM), matrix-pattern MHKS, and matrix-pattern semi-supervised LapMatLSSVM (Laplacian matrix-based least square SVM), the experiments on forest fire images verify that our proposed algorithm has higher fire image recognition rate and less training time.

Key words:  forest fire recognition, vector-pattern, matrix-pattern, bilinear function, semi-supervised learning