Deep Learning Based Hazard Label Object Detection for Lithium-ion Batteries Using Synthetic and Real Data

Changing transport regulations for intralogistics tasks leads to the need for object detection of hazard labels on parcels in high-resolution grayscale images. For this reason, this paper compares different Convolutional Neural Network (CNN) based object detection systems. Specifically, a YOLO implementation known as Darkflow as well as a self-developed Object Detection Pipeline (ODP) based on the Inception V3 model is considered. Different datasets consisting of synthetic and real images are created to set up the necessary training and evaluation environments. To check the robustness of the systems under real operation conditions, they are assessed by the mean Average Precision (mAP) metric. Moreover, results are evaluated to answer various questions like the impact of synthetic data during training, or the highest quality level of the systems. The YOLO models showed a higher mAP and a much higher detection speed than the MSER Object Detection Pipeline at the cost of higher training times. The mixed training data set with synthetic and real data showed a slightly reduced mAP on the validation set compared to just real data.