Smart detection of seepage in river dikes based on thermal infrared images

摘要: In this thesis an algorithm is developed to detect seepage in river levees based on thermal infrared images. From previous research it known that contrast (a cold levee body and warmer water leaking through the for example) can be a good, fast non-destructive indicator of which may eventually lead failure. Early detection essential good management practical tool simplify help guards keeping what protects safe. A supervised learning approach chosen allows continuously improve new training testing data. Such feasible because image acquisition itself still needs improved only few data was available at time research. first step, suspicious areas (blobs) are detected image. To do so pre-processed contours derived using maximally stable extremal regions. As result step collection blobs becomes available. Using simple post-processing some removed without much computational effort. second set features represent each these blobs. Three kinds implemented. kind temperature measurements characteristics. These used feed classifier separates leaks from objects, imperfections other not correspond leak. decide best used, recursive feature elimination guided by cross validation It inherent problem there imbalance between number actually leak not. This impedes classification task hand. deal with this, different sampling techniques Next set, diversity true examples Therefore, artificial samples generated 1. third support vector linear kernel implemented perform actual classification. has disadvantage allowing relatively decision boundaries but advantage less likely over-fitting Excellent results were obtained slightly accurate test while classifier.