Object Recognition in Underwater Imaging Using Machine Learning Techniques. - Venkataraman Padmaja

. .

Résumé :
For many years, mines in the ocean are becoming major worry and threat to human lives and vessel safety. These mines are generally placed in the ocean for security reasons to protect from enemies which can destroy submarines and ship which comes in contact with the mines. It's very difficult to identify and detect the objects in underwater using sonar imagery because of its complications. This is due to factors which involve variations in operational and environment conditions, spatially variable chaos, variation in target shapes, structure and orientation. Considering all these conditions, a method had been proposed which can detect and classify whether the object is a mine or an object which resembles a mine under water. Images are obtained from sonar camera scanner, which is placed in underwater communication network in a moving vehicle with a sensor. In our application, using image processing and machine learning technologies, we have studied the behaviour and differentiate the features of a mines and rocks. In most cases mines are considered to be metal objects. We have designed this application with algorithm which can give a real time capability to detect the objects and distinguish them as seabed objects and imaginary artifacts which are induced by vehicles. UCI dataset is considered in this system which holds all possible attack data with percentage of possibility. This data is then utilized in the pre-processing by applying one hot encoding. Statistical methods are used to obtain Z-scores, mean, median and mode for determining the significant features and training on 80% of dataset is validated. The remaining 20% dataset is tested and validated using Decision tree, K-NN and Gradient boosting algorithm. The efficiency of these algorithms is being analysed and discussed and it is found that Gradient boosting algorithm is best suitable algorithm to be utilized for development of mine detection. Different parameters were interacted based on the testing and checked for false negative rate, accuracy, f-score, precision time. These metrics are vulnerable to the efficiency of the mine detection model that is being proposed.