The Image of Numerals with Calibration on TVM

Today's open-source deep learning frameworks, such as TensorFlow, PyTorch, TVM, and Caffe, have become widely used tools for deep learning and support hardware accel-erators like GPU and TPU. Among these platforms, TVM stands out for its flexibility and compatibility by compiling high-level deep learning frameworks (such as TensorFlow, PyTorch, etc.) to generate optimized low-level code, which can be automatically optimized to improve performance. Although CNN significantly affects image recognition, some images may not be correctly recognized due to problems such as location, orientation, and deformation. Therefore, this paper aims to improve recognition accuracy by normalizing images. This paper focuses on developing an image normalization calibration module on TVM with OpenCV and performing recognition for digital photos. This module is mainly responsible for edge detection and performing normalization adjustments based on the selected preprocessing areas.