Deep Learning (DL) techniques for Image Analysis have been shown to be highly effective across a broad range of applications including inspections, health screenings, image indexing, medical imaging, security, and surveillance.
DL solution for a given task are developed in 6 steps:
◆ Collection and selection of representative data.
◆ Manual annotation of training data to associate the known state or properties to each element.
◆ Training database generation.
◆ Selection of the most suitable DL network architecture for solving the problem and returning the states or properties of the elements (image or object classes, object bounding boxes, etc.) in the data to be studied.
◆ Training of the network using the annotated data to adjust its weights until the computed outputs are as close as possible to the ground truth
◆ Evaluation of the model using a subset of the database not used for the training.
These DL techniques could not have been developed without a large training database, necessary for achieving state-of-the-art results. So ADCIS has developed a software package:
The graphical user interface is the key to speeding up the annotation by reducing mouse clicks, having the proper drawing tool to outline the objects, providing shortcuts to select the annotations to perform, etc. Once configured, the software matches perfectly the user’s needs and allows her/him to annotate in the best conditions. In addition, Annotate can pre-annotate images by applying a DL network already generated. The user only needs to check the annotations and edit them if required. This can speed up significantly the annotation process. The training database is built up through the annotation process and the user is informed when all the images and video frames are annotated.