Enhanced Fire Detection Algorithm for Chemical Plants Using Modified YOLOv9 Architecture

In the high-risk environment of chemical plants, the significance of fire detection cannot be overstated, placing stringent requirements on an algorithm's real-time responsiveness, accuracy, and resistance to interference. This study heralds the inaugural application of the YOLOv9 algorithmic model for the task of flame detection within such facilities, introducing an innovative modified YOLOv9 model developed explicitly to enhance precision and expedite detection rates in the crucial domain of chemical fire surveillance. Our proposed methodology is underpinned by an extensive, self-compiled dataset of fire instances, enriched with data augmentation techniques aimed at bolstering the model's operational efficacy. We incorporate an array of visually similar fire-like images and genuine flame photographs, thereby establishing a robust database purposed for flame detection. Significant advancements have been made to the YOLOv9 architecture, streamlining the original network, accompanied by the adaptation of the bounding box loss function to Focaler-IoU, augmenting the model's delineation accuracy. The experimental results are promising; the optimized YOLOv9 model has achieved a computational complexity and model weight of 187.1GFLOPs and 85.7 MB, respectively, which represents a reduction of 26% and 17% when juxtaposed with the archetype. Nevertheless, the model maintains commendable detection precision with an accuracy rate of 94.9% and an mAP50 value peaking at 96.1%, virtually paralleling the baseline model. These enhancements afford a more compact model size without compromising on detection capabilities, adequately fulfilling the fundamental flame detection requirements necessary for the challenging ambiance of chemical manufacturing environments.