Reducing False Alarms in Lightweight Indoor Fire Detection

الملخص

Deep learning fire detectors are almost always judged by mean average precision on a test set of fire images. That number says nothing about how often the model raises an alarm on a lamp, a candle, or a sunset. We trained six small YOLO detectors on an indoor fire dataset and found that every one of them fired on a quarter to a third of ordinary fire-like images, even though their detection scores looked healthy. This paper describes how we brought that rate down for YOLO26n, the smallest of the six and the model we deploy. We added background images with no labels to the training set, first chosen by hand and then mined automatically from a large open image collection using CLIP features and clustering. On 500 held out images of lamps, bulbs and candles the false alarm rate fell from 27.0 to 0.4 percent once both sources were combined, and on 196 images from fire-like categories that appear nowhere in training it fell from 19.4 to 1.2 percent. Distilling a larger teacher into the deployed model then recovered the accuracy given up along the way, reaching 0.9325 mAP50 with 2.5 million parameters and 5.8 GFLOPs.