Research On Image Recognition Algorithms In Cell Screening And Fire Detection

Image recognition mainly includes image preprocessing,feature extraction,classification.Digging deep into the feature of the image and improving the reliability of feature extraction is the key of image recognition.Image recognition can be divided into static target recognition and dynamic object recognition.Static target recognition can be accomplished by single frame image analysis,while moving object recognition is usually based on the temporal and spatial variation of time and space of the video or image sequences.Combining characteristics of the image and relevant knowledge of the object,to design an image recognition system with higher recognition rate and better real-time performance is the final destination of image recognition.Cervical cells identification and fire smoke detection are two hot issues in academic circles today,therefore this paper selects them as the research objects.The main work is summarized as follows:For complex static targets,image segmentation is difficult and time-consuming.In order to solve this problem,this paper deeply analyzes the characteristics of H&E stained cervical cells.The texture and color characteristics of different blocks are found to be of great difference.Therefore,the idea of using block image processing instead of image segmentation is proposed.A cervical cell screening system based on block image characteristic analysis is designed.Firstly,the background blocks are removed by using the characteristic difference between the background and the non-background blocks to reduce the subsequent processing time;Then,the non-background blocks are divided into normal blocks and abnormal blocks,and the characteristics of these two kinds of blocks are analyzed deeply.11 features with significant differences(with t-test,p<0.001)are extracted;Finally,the structure and parameters of support vector machine(SVM)are designed.According to the contrast experiments,compared to the recognition system based on segmentation,the real-time performance proposed in this paper is greatly improved on the basis of ensuring the accuracy.Conventional image recognition system usually requires manual characteristics,making the system has a strong limitation.In terms of cervical cells,the existingrecognition system usually works well on one kind of staining cells,and for different staining cells,different identification systems are needed.To solve this problem,the characteristic learning ability of convolutional neural networks(CNNs)is combined with the classification advantages of(SVM.The network is constructed with SVM replacing the traditional CNNs output layer,and parameters are designed according to the features of the cells.Experimental results show that the proposed recognition system outperforms the CNNs and SVM.Static features cannot describe the dynamic characteristics of the target.Therefore,a combination of static and dynamic features is required to detect the moving target.In order to detect the fire early,the the contour optical flow vector of the suspicious area is extracted,instead of the whole suspicious area.A set of‘dynamic-static-dynamic' identification method is proposed: Firstly,the moving object is detected by the improved frame difference method,so the interference of static object is excluded;Then,by detecting the color model and texture feature of the moving area,some motion disturbances are excluded;Finally,based on the Gauss Pyramid Lucas-Kanade optical flow method,the optical flow vector of the contour of the moving region is analyzed,so the non-smoke interference is excluded.The comparative experimental results show that the real-time performance proposed in this paper is improved on the basis of ensuring the accuracy.