The objective of the thesis is to implement a system for semi-automatic counting of polyps on the photographs of sea-bed. Numerous photographs have been obtained, wherein the manually labelled sample of those was used for the train set, and on basis of this the machine learning was implemented. Photographs had to be re-processed with appropriate methods in order to correct the bad lighting and furthermore even out the differences. First a model was used, which removes the background in the image while leaving everything that is part of the polyp. Image with the removed background is re-processed again for the model for automatic labelling of polyps. The user can then decide whether automatically labelled polyps will be repaired and thereby the exact number of polyps is obtained, or leave it to the model for predicting the number of polyps. This model is taught to foresee the error of previous models and to predict the number of polyps more accurately. The entire system was tested by test images and it was established that some images are not suitable for the entire system, and, therefore cannot be used in this system. While manually correcting automatic labels of polyps it was established that it is significantly faster than the manual labelling of polyps in the overall photographs. The model for predicting polyps was also positively assessed, as it reduces the error of the model for automatic labelling of polyps on average.
|
