Early Fire Detection Based On Improved YOLO-PCA Network

Fire occurs in nature and is beneficial for humans because of its usage on domestic and commercial levels.Along with its inevitable existence,the fire incidents cause disastrous impacts on a country’s economy,human life,and the surrounding zones.Preventive precautions and measures are foremost to ensure fire safety.Early detection of fire before its growth to the critical stage is crucial.Among the various fire signatures,smoke is one of the most prominent signs because it usually evolves earlier than flame.Early detection of smoke having indistinguishable pixel intensities is a difficult task.To solve the problem,we have integrated the principal component analysis(PCA)as a preprocessing module with the improved version of You Only Look Once(YOLOv3).This work also presents the superiority of computer vision in fire detection systems over the traditional contact-based detectors.The core research work and conclusions are summarized as follows:1.The ordinary YOLOv3 structure has 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 lower spatial information feature m aps and fine-grained semantic information in upsampled maps via skip connections and concatenation operation s.In addition,for the processing of immense small smoke images as positive samples during training,new sizes of anchors have been calculated on small smoke data at a specified Intersection over Union(IoU)threshold.A relatively new K++means clustering approach clusters the input data points and generates the respective dimensions of the boxes.2.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 pixels in the testing of the trained network hence,overcoming the latency at the inference stage.To reduce the dimensional complexity present in the smoke datasets,we have integrated an unsupervised dimen sion reduction module called Principal Component Analysis(PCA)at the input of the improved YOLOv3.The input smoke and non-smoke images are pre-processed before feeding the network.3.This work optimizes the training scheme for YOLOv3.The training scheme of the improved network and the traditional network is the transfer learning from the backbone of YOLOv3,called Darknet-53.For the training of the suggested network,we fine-tune the fewer initial layers and freeze the final classifying layers of a source network.The proposed training scheme adopts a specific learning pattern instead of random weight initialization.4.The experimental results show the improvements in precision rate,recall rate,and mean harmonic(Fl-score)by 2.67,3.06,and 5.59 percentages.The respective improvements in average precision(AP)and mean average precision(mAP)is 1.66 and 2.78 percentages.