Research On Local Feature Learning And Representation For Single Image Smoke Recognition

Smoke detection provides important assistances for early fire alarming.Computer vision based smoke detection has wider working regions,thus it becomes a crucial research topic of fire alarm.Single image smoke recognition is the basic task of video smoke detection,and smoke feature representating and statistical recognition model learning are key steps of smoke recognition.Effectively representing the intrinsic structures of smoke helps increase the accuracy and confidence of smoke recognition.Consequently,the application of computer vision techniques in fire detection and alarming can be driven,and the development of recognition model learning and feature representation algorithms can be promoted.The colors,shapes and movements of smoke are susceptible to burning objects and combustion environment,so it's difficult to describe smoke.Thus not only makes image-based smoke features lack discriminative ability and robustness,but also leads to high false alarm rates.To this end,exploring feature representing methods and recognition models that can preserve intrinsic smoke component is of great significance.On one hand,local features can effectively represent texture,which is the most robust attribute of smoke.Traditional local feature representing methods suffer from three obstacles:(1)they need manual intervention,(2)they are applied to single images so the representation ability is limited,(3)most discriminative components,such as high order,multi-scale and transform invariant ones,are simply concatenated,thus leads to big computation consumption,high redundancy and low efficiency.Deep learning-based frameworks avoid the above mentioned shortcomings but lack interpretability and flexibility.In summary,existing algorithms fail to balance between representation ability,intelligence and flexibility.On the other hand,generative recognition models describe the relationship between input features and output labels and preserve intrinsic structure of smoke.However,they are under-utilized in smoke recognition task.This paper studies on key techniques for smoke feature representation.First,existing problems are concluded through reviewing literatures,then learning steps are involved as solutions to the problems so that algorithms can automatically select the most intrinsic and stable component to represent smoke.In addition,a new smoke recognition framework is proposed.Thus accuracy of smoke recognition can be increased by the proposed feature representing methods and recognition frameworks.Eventually,the practicability and industrialization of smoke detection and early fire alarming may be pushed forward.The research contents of this paper are summarized into four parts.The former three provide different solutions to improve feature representation performance by involving feature learning.(1)Learning-based high-order smoke feature representation.Learning based strategy and feature transformation methods are studied.First,patch-based local differences are computed.Then the sampling strategy is learnt to increase the compactness of basic features and to surpress noise sensitivity.Third,the sampled differences are used to learn a model,through which middle-level features that are more seperable can be obtained.At last,the features are encoded without quantization to avoid too much information loss.(2)Multi-scale and multi-order smoke feature learning and representation.Middle-level and high-level feature representing methods are studied,and crossscale and versatile informations are extracted by involving learning step.First,we construct a scale space and densely slide a 3D sampling window in the space to compute 3D local differences across scales.Then,we holistically learn a projection model from all 3D local differences of training images.The learnt model is used to generate middlelevel features,which are more separable.Third,we process these feature maps in within-and between-map encoding ways.The former captures local texture information while the latter models texture distributions across feature maps.Thus,multi-order representation for middle-level features is achieved.The above steps can be stacked on top of each other to present a hierarchical structure,through which multi-order,hierarchical,robust and discriminative features can be obtained.(3)Multi-orientation,multi-scale and invariant feature learning and represention.Learning-based methods for stable component extracting methods are explored and solutions to problems existing in multi-orientation and multi-scale feature representations are studies.First,multi-scale and multi-orientation local response maps are generated by Gabor filters.Then condensing methods along different scales and orientations are proposed to preserve invariant components.At last,we leverage within and between-channel encoding methods to capture rotation and illumination invariant features.In addition,we propose two extensive representation ways for features in the shallow layer.One is to apply 3D convolution based holistic learning to increase feature separability.The other one restores and encodes the indices of max responses to preserve stable global texture distribution.The above steps can be stacked to form a hierarchical structure,termed Gabor Net.The proposed Gabor Net provides multi-scale,multi-orientation and invariant features representation.(4)A generative model based high efficient smoke recognition framework.We study in the principle of generative models represented by GPR and combine non-linear projection methods with GPR to present a new smoke recognition framework.The linear correlation between smoke features are removed while lowdimensional manifold structure embedded in smoke is preserved.As a result,the speed of smoke recognition is improved and false alarm rates are decreased.