Research And Implementation Of Weather Recognition Based On Machine Vision

Machine vision is the technology and methods used to provide imaging-basedautomatic inspection and analysis for such applications as automatic inspection,process control, and robot guidance in industry. In recent years, with the continuousimprovement of computer hardware capabilities, machine vision technology can beextended to more applications, such as face recognition, unmanned vehicles, roadtraffic management, product quality and grade classification. Weather conditions as anatural phenomenon is out of our control, have an impact on people’s life. Althoughtoday, more and more intelligent devices can replace the human to complete some ofthe highly difficult task, but there are still some application scenarios intelligentsystems need to do some judgments and choices depend on weather condition. Thestudy how to recognize the weather that intelligent systems around has practicalsignificance. The main technology and method are image process and patternrecognition.Feature extraction is one of the most important module about pattern recognition.It provides original pattern vectors input to classifier. The quality of extracted featureshas enormous influence on designing classifier. We always want the feature extractedcontain large amount of classification information and to extract as more feature aswe can. This paper extracted the contrast, sharpness, power spectrum slope, textureand color features. And, we also proposed a more efficient way to extract thosefeatures in the way which divide the whole picture into lots of sub-image. It isconvenient to eliminate the area where has little classification information in that way.After get the features, we also uses the K-W check, ROC, and PCA to reduce thedimension of the feature vector. Then, we design a classifier based on SVM with RBFkernel function and use cross validation to find the optimal parameter of the SVMmodel and kernel function. And also we design an ADAG decision tree which used tosolve multi-class problem of SVM.At last, we programed all algorithm and test them on Matlab with the sampleimage set of Columbia University WILD and got a high efficiency and accurate. The test data are mainly the WILD image databases. Test results show that this system cancompletely accomplish the weather phenomena recognition task.