Image-based Detection And Counting Of Soot Particles

Soot is a kind of inhalable particulate which contains many toxic substances.It mainly comes from incomplete combustion of diesel and coal.Its particle size is uncertain and generally appears in the form of agglomeration.Collecting data of particle size and number of soot will help researchers understand its properties,thereby optimizing the combustion process of coal combustion and reducing the amount of toxic particles in diesel exhaust.The soot images obtained under the transmission electron microscope are generally highly stacked,and the edges of the particles are blurred.It is time-consuming and laborious to use traditional manual detection method,which cannot meet cutting-edge scientific research requirements.In recent years,the development of digital image processing and convolutional neural networks has made great progress,also great breakthroughs in imagebased object segmentation,image recognition,target number estimation and other fields.Therefore,This paper takes the lead in putting forward the automatic detection and counting methods of soot particles using these two technologies,and carries out the following research:For the task of detecting soot particles,this paper proposes an improvement on the basis of image segmentation based on fuzzy C-means clustering.Different from traditional pixel-level fuzzy clustering,we use superpixel segmentation technology to segment the image foreground into small regions with similar grayscale,and then fuzzy clustering on the basis of region-level.In feature selection,we not only use regional grayscale information,but also use location information to ensure that clustering results are in close to each other.In the selection of the optimal cluster number,we introduce the normalized uniformity measure(UN)to evaluate the quality of segmentation,and guide the clustering algorithm to select the best number of clusters.Finally,by comparing with the traditional Hough circle detection algorithm,the paper analyzes the difference between the proposed method and the traditional method from the perspective of subjective evaluation and objective calculation.The results show that the soot particles detected by our method have fewer false detections and missed detections,and higher uniformity measures.For the task of counting soot particles,we start with the basic data preparation,and introduce in detail the production process of the data-set used in the experiment.Because of the large amount and serious aggregation of soot particle,we propose a neural network framework called Deep Parallel Network(DPN)to automatically extract soot image features and estimate its density map.The parallel network perceives different scales of soot particles through multiple different receptive fields so that the extracted features are more accurate.Our model outputs density maps instead of direct numerical estimation,which will make our network training less difficulty and get more information about particle distribution.Finally,we design several sets of comparative experiments from different angles,and the experimental results show that our method has higher accuracy and robustness,and the average prediction error of the number of particles in test images is less than 8%.