小样本火灾图像色彩关联的模糊隶属度识别方法

A fuzzy membership recognition method for color correlation of small sample fire images

针对火灾图像识别过程中样本数据需求量大、计算复杂度高等造成的识别效果不理想问题,提出一种小样本火灾图像色彩关联的模糊隶属度识别方法。首先,计算待检图像和常规图像的偏色因子大小,同时进行罗曼洛夫斯基准则优化,在对偏色因子向量由小到大排序的基础之上确定待检图像类别;其次,基于灰色相对关联度和接近关联度以及权重求解的基础之上计算灰色综合关联度,同时求解待检图像对于同类别火灾图像的隶属度大小;最后,基于最大隶属度原则与图像类别阈值,判断图像是否属于火灾图像。结果表明:(1)该方法对于火灾图像的综合识别精度达88.89%,同时对非火灾图像的识别精度达100%。(2)在室外晴天条件下,自然光对蓝色背景的火灾图像色彩特征具有一定干扰性;在室外阴天条件下,自然光对绿色背景的火灾图像色彩特征具有一定干扰性;但上述两种场景下自然光对红色背景的火灾图像颜色特征不具有干扰性,识别精度达100%。(3)在暗箱无光条件下,对于红绿蓝三种背景的火灾图像综合识别精度均达100%,这表明自然光对火灾图像的颜色特征具有一定干扰性。(4)不同颜色背景的火灾图像的灰色综合关联度具有明显差异性;对于红色和蓝色背景火灾图像,相对关联度的贡献度明显高于接近关联度;而对于绿色背景火灾图像,接近关联度贡献度明显高于相对关联度。(5)与决策树、支持向量机、深度神经网络算法相比,该方法针对小样本图像具有较高的识别精度,以及较小的计算成本。

A small-sample fuzzy membership recognition method based on color correlation of fire image is proposed to address the problem of unsatisfactory recognition performance caused by the large sample data requirements and high computational complexity in fire image recognition.First,the size of the color deviation factor between the target image and the reference images is calculated,and the Romanovsky criterion optimization is performed to determine the category of the target image based on the sorting of the color deviation factor vector from small to large.Second,the gray comprehensive correlation degree is calculated based on the relative correlation degree,the proximity correlation degree,and the weight solving,and the membership degree of the target image to the same category of fire images is determined.Finally,based on the maximum membership principle and the image category threshold,the image is judged whether it belongs to a fire image.The results show that the proposed method achieves an overall recognition accuracy of 88.89%for fire images and 100%for non-fire images.Under sunny outdoor conditions,natural light interferes with the color characteristics of fire images with blue backgrounds.Under outdoor cloudy conditions,natural light interferes with the color characteristics of fire images with green backgrounds.However,natural light does not interfere with the color characteristics of fire images with red backgrounds in the above two scenarios,achieving 100%recognition accuracy.Under dark box conditions without light,the overall recognition accuracy for fire images with red,green,and blue backgrounds all reach 100%,indicating that natural light interferes with the color characteristics of fire images.The gray comprehensive correlation degree of fire images with different color backgrounds shows noticeable differences;for fire images with red and blue backgrounds,the contribution of relative correlation degree is significantly higher than that of proximity correlation degree;while for fire images with green backgrounds,the contribution of proximity correlation degree is significantly higher than that of relative correlation degree.The proposed method has higher recognition accuracy and lower computational cost for small-sample images compared with decision trees,support vector machines,and deep neural network algorithms.