图像监测是矿井火灾火焰主要感知方法，但受矿井光源影响。内凹度能减小摄像机的安装位置、拍摄距离和拍摄角度的影响，排除矿井光源干扰，快速识别并排除矩形光源等非弧形类光源正视图及其变形图像；但对于圆形光源和类圆形光源等弧形类干扰光源，计算量较大，识别时间较长。圆形度能够排除圆形光源的干扰，但难以排除非圆形光源的干扰。矩形度能排除矩形光源的干扰，但难以排除非矩形光源的干扰。受摄像机安装位置和拍摄角度等影响，圆形和矩形光源图像会出现变形，无法呈现理想的规则形状。因此，圆形度和矩形度算法难以排除矿井光源的干扰。揭示了火焰图像边界上存在多个凹陷区域，总凹陷区域面积较大；圆形灯、长方形灯和正方形灯等矿井实际光源图像边界上不存在凹陷区域等特点。提出矿井外因火灾图像凹陷度识别方法，计算目标图像边界总凹陷区域面积与其目标图像实际面积的比值（即图像凹陷度），根据火焰图像凹陷度数值较大，而矿井光源图像凹陷度数值较小，区分矿井光源与火焰。研究提出的凹陷度方法不受摄像机安装位置、拍摄距离和拍摄角度、矿井光源形状和大小等影响，适应性强，识别准确率高。凹陷度识别方法计算得到的矿井干扰光源图像与火焰图像凹陷度平均差值绝对值最大，波动较小，区分度最好。内凹度识别方法计算得到的矿井干扰光源图像与火焰图像内凹度平均差值绝对值较大，波动较小，区分度较好。矩形度识别方法计算得到的矿井干扰光源减去火焰图像矩形度平均差值绝对值较小，波动较大，区分度较差。圆形度识别方法计算得到的矿井干扰光源减去火焰图像圆形度平均差值绝对值最小，波动最大，区分度最差。结果表明：凹陷度识别火灾火焰图像准确率为98.2%，召回率为98.4%，准确率和召回率最好；内凹度识别火灾火焰图像准确率为92.8%，召回率为92.4%，准确率和召回率较好；矩形度识别准确率为86.5%，召回率为86.5%，准确率和召回率较差；圆形度识别准确率仅为35.9%，召回率为31.9%，准确率和召回率最差。

Image monitoring is the main perception method for mine fire flames, but it is affected by mine light sources. The internal concavity can reduce the influence of camera installation position, shooting distance, and shooting angle, eliminate interference from mine light sources, and quickly identify and eliminate the front view and distorted images of non arc-shaped light sources such as rectangular light sources; However, for arc-shaped interference light sources such as circular light sources and quasi circular light sources, the computational complexity is large and the recognition time is long. Circularity can eliminate interference from circular light sources, but it is difficult to eliminate interference from non-circular light sources. Rectangularity can eliminate interference from rectangular light sources, but it is difficult to eliminate interference from non rectangular light sources. Due to factors such as camera installation position and angle, circular and rectangular light source images may become distorted and unable to present ideal regular shapes. Therefore, it is difficult to eliminate the interference of mine light sources using circularity and rectangularity algorithms. It reveals that there are multiple depression areas on the boundary of the flame image, with a relatively large total depression area. However, there are no depression areas on the boundary of actual mine light source images such as circular lights, rectangular lights, and square lights. Propose a method based on the recognition of depression degree in images of external mine fires, calculating the ratio of the total concavity area of the target image boundary to the actual area of the target image (i.e. image depression degree). Based on the larger depression degree value in flame images and the smaller depression degree value in mine light source images, distinguish between mine light sources and flames. The depression degree method proposed in this article is not affected by the distance and image size between the camera and the detection target, the installation position and angle of the camera shooting the detection target, the shape of the mine light source, etc. It has strong adaptability and high recognition accuracy. The average difference in depression degree between the mine interference light source image and the flame image calculated by the depression degree recognition method has the largest absolute value, small fluctuations, and the best discrimination. The average difference in the internal concavity between the mine interference light source image and the flame image calculated by the internal concavity recognition method is relatively large in absolute value, with small fluctuations and good discrimination. The rectangular degree recognition method calculates that the absolute value of the average difference between the rectangular degree of the mine interference light source and the flame image is in the middle, with large fluctuations and good discrimination. The roundness recognition method calculates the minimum absolute value of the average difference between the roundness of the mine interference light source and the flame image, with the largest fluctuation and the worst discrimination. Experimental studies have shown that the accuracy of recognizing fire flame images based on depression degree is 98.2%, and the recall rate is 98.4%, the best accuracy and recall; The accuracy of identifying fire flame images based on the internal concavity is 92.8%, and the recall rate is 92.4%, the better accuracy and recall; The accuracy of rectangularity recognition is 86.5%, and the recall rate is 86.5%, the worse accuracy and recall; The accuracy of circularity recognition is only 35.9%, and the recall rate is 31.9%, the worst accuracy and recall.