Comparison of CNN for the detection of small ojects based on the example of components on an assembly table

Abstract

This paper presents the results of a comparison of deep neural networks for detection of small objects typical for manual manufacturing tasks. We created a set of training, validation and evaluation data and selected four state of the art deep neural networks for object detection. We trained them with the same number of epochs, 200 epochs per network architecture and compared the training time, accuracy and prediction time on evaluation data. Additional we compared the neural networks on thirty images of three very small and similar components.