Research On Vision Detection Method For Forest Fire Smoke Visible Image

Forest fire is sudden and destructive,which not only threatens the safety of people and property,but also damages ecological environments.Therefore,early forest fire automatic detection and timely warning is important to protect forest resources and reduce disaster damages.Currently,visible light based computer vision detection technology is the main method to monitor forest fires.However,the actual forest environment is complex,and smoke images are susceptible to a variety of changing factors such as geographical location,seasonal alternation,cloudy weather,shooting angles,and smoke characteristics.To solve the above problems,this thesis focuses on similarity factors such as clouds,fog and haze in forest natural environment,monitoring distance and monitoring timing factors as well as forest topographic environment diversity factors,and research on early forest fire smoke detection from two aspects: supervised learning and unsupervised learning.The main innovative work can be summarized as follows:1.A domain adversarial feature fusion network for forest fire smoke detection is proposed.Firstly,faced with the complex background interference in forest environments,a dual-channel feature fusion network which contains a densely dilated convolutional neural network and an attention-based network with skip connection is proposed to extract discriminative smoke features.Specifically,the densely dilated convolutional neural network is used to extract deep and abstract features,while attention-based network with skip connection is used to extract shallow and detailed features.To address the problem of domain bias arising from differences domains,an adversarial feature adaptive network is introduced to improve the generalization capability.Furthermore,to tackle the problem of simultaneous optimization of label classifier and domain discriminator,a joint optimization strategy for domain adversarial feature fusion network is designed to increase the transfer gain of our method.Finally,the effectiveness and generalization ability of the domain adversarial feature fusion network in excluding the interferences from natural environmental and forest topographic diversity is verified on two self-built forest fire smoke datasets and three publicly available fire smoke datasets.2.An attention-based prototypical network for forest fire smoke few-shot detection is proposed.To address the problem of capturing tiny forest fire smoke at a long distance,a shallow neural network based feature extraction module is designed to retain global information of tiny smoke.Additionally,to address the information missing problem caused by few pixels of smoke,an attention module is introduced to focus on small objects.Then,an attention map is generated for each feature by correlation estimation and meta-fusion to highlight the target object and improve the discriminability of image features.To address the overfitting problem caused by training few-shot samples,a meta-learning module based on Euclidean distance is designed which classify each query sample via the distance of query feature and class prototypes.Finally,the effectiveness of the attention-based prototypical network in excluding the factors of monitoring distance,time and forest terrain diversity is verified on forest fire smoke few-shot dataset and mini Image Net dataset.3.A self-supervised contrastive learning network for small smoke detection in forest fire video is proposed.Firstly,to retain both local features and global information of smoke images,a cross-dual network based on convolutional neural network and visual Transformer structure is proposed to enhance the feature representation of tiny smoke.Then,due to the limited quantity of tiny smoke images,a contrastive learning method of unsupervised self-distillation network is introduced to generate pseudo-labels and learn the semantic information of unlabeled samples.Furthermore,to address redundancy of contextual information in continuous video frames,a motion region extraction module is designed to process attentional feature maps to identify motion targets in area of interest for tiny smoke detection and localization.The effectiveness and stability of the contrastive self-supervised learning network in excluding natural environmental disturbance factors of monitoring distance,time and forest topographic environment diversity is verified on both self-built forest fire tiny smoke video dataset and publicly available forest fire smoke video dataset.Finally,the performance of the supervised and unsupervised learning forest fire smoke detection methods is compared and analysed in terms of smoke dataset type,smoke feature extraction network structure and unsupervised and supervised learning tasks,respectively.