Early Smoke Detection Based on Improved YOLO-PCA Network

Abstract

Early detection of smoke having indistinguishable pixel intensities in digital images is a difficult task. To better maintain fire surveillance, early smoke detection is crucial. To solve the problem, we have integrated the principal component analysis (PCA) as a pre-processing module with the improved version of You Only Look Once (YOLOv3). The ordinary YOLOv3 structure has been improved after inserting one extra detection scale at stride-4 specifically to detect immense small smoke instances in the wild. The improved network design establishes a sequential relation between feature maps of lower spatial information and fine-grained semantic information in up-sampled maps via skip connections and concatenation operations. The testing of the improved model is carried out on self-prepared smoke datasets. In digital images, the smoke instances are captured in various complicated environments, for example, the mountains and fog in the background. A principal component analysis (PCA) helps in useful features selection and abandons the involvement of redundant features in the testing of the trained network hence, overcoming the latency at inference stage. In addition, to process small smoke images as positive samples during training, new sizes of anchors are calculated on small smoke data at a specified Intersection over Union (IoU) threshold. The experimental results show the improvement in precision rate, recall rate, and mean harmonic (F1-score) by 2.67, 3.06, and 5.59 percentages. The respective improvements in average precision (AP) and mean average precision (mAP) are 1.66 and 2.78 percentages.