Deep Learning in Semiconductor Devices for Road Damage Inspection Applications

The advances in semiconductor technologies are enabling popularity of deep learning for various applications including infrastructure inspection. The road damage inspection is time consuming, if done manually. If deep leaning techniques are applied on collected road images, it can significantly reduce time and cost in inspections. It may be adequate to upload collected road images to cloud for deep learning inferencing, if there is sufficient link bandwidth from drone or inspection vehicle. If not, it is beneficial to perform first level deep learning inferencing on drone or inspection vehicle and upload images of interest to cloud for detailed inferencing. In this paper, hierarchical deep learning inferencing method has been proposed using semiconductor edge devices, such as ASIC or FPGA on drone or inspection vehicle. Also presented deep learning training system that runs on cloud. Further, FPGA based platform has been deployed, which achieved less than 0.4 seconds inferencing time per frame.