A Detector for Stacking Objects Based on Neural Network

This paper aims to design a detector based on a neural network to detect workpieces in a multi-stacking situation. Because our samples have similarity with each other which will make our model overfitting without any data enhancement method. We propose a novel way to enhance our data, which is that adding "stack" conditions to our single sample and use transfer-learning at the stage of training which is a normal way to solve the problem of training network with the tiny dataset. Our sampling method partly eliminates the correlation between samples to some extent and the model trained on our dataset performs well. As a result, we get the accuracy of classification with 96.04% and the mAP of object-detection with 91.12% and 92.61%. This paper gives a new sampling method to solve the problem of detecting multi-stacking objects. For details, this paper will introduce from the method for training data sampling to the performance of object-detection.