وحيد سعيدي مجد

By deploying deep learning models, computers have achieved great results in object detection. Deep learning models are needed to be trained by a huge dataset to achieve the ability for generalizing their knowledge to application domain. Preparing datasets is costly and time consuming. Using active learning, we try to select the most valuable data, such that the model learn best representation and knowledge of application domain. In other words, we try to allocate our limited resource for gathering datasets. We try to expend minimal time and cost to train saffron flower model by active learning. In active learning, data instances are measured by uncertainty and diversity criterion. We eva‎luated least confidence sampling, a method for measuring uncertainty criterion, for saffron flower detection in image level and object level in experiments. Furthermore, we proposed weighted random sampling method to measure uncertainty and diversity criterion, jointly. By experimenting this method, we managed to achieve higher optimality in accuracy and time. We experimented active learning in object level mode with square patches in two different sizes, one 100 pixels long and the other 50 pixels long. In both sizes, the non-uniform random sampling method had the best results; It achieved $ 0.8617$ mAP at 100 pixels and $0.8582$ mAP at 50 pixels. The number of foreground tags in the same method was 1662 in the size of 50 pixels and 2387 in the size of 100 pixels. Since the number of labels representing the cost is much smaller at 50 pixels than at 100 pixels, we conclude that the non-uniform random sampling method on the surface of an object with a length of 50 pixels is the better method.

يادگيري فعال , يادگيري عميق , تشخيص اشياء , تشخيص گل زعفران

Active learning , Deep learning , Object detection , Saffron flower detection

Vahid Saeedi Majd

Mohammad Reza Mohammadi