Quantity Estimation Method of Warehousing Cargo Based on CSwin-Unet and Pixel Position Weight
Jiehao Wu, Guangyuan Zhang, Kefeng Li, Peng Wang, Zhikang Li · 2022 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML) · 2022
The quantity estimation of cargo is of great significance for the scientific management of storage. The traditional quantity statistics of cargo is based on RFID technology to obtain the information of cargo in circulation, which is time-consuming, laborious, and complicated. Due to the unsatisfactory accuracy, computer vision-based statistical methods are not widely used in the field of cargo quantity statistics. To solve the above problems, this paper proposes a method for estimating the quantity of cargo in storage based on CSwin-Unet and XGBoost, which improves the accuracy while reducing equipment dependency. Firstly, the CSwin-Unet model is pre-trained on the COCO Stuff dataset using transfer learning. It achieves 48.8% mIoU, which is 1.6% higher than Swin-Unet. Secondly, this paper designs a pixel position weight (PPW) loss function to fine tune the CSwin-Unet in local dataset to obtain better feature extraction effect. Estimation of the cargo quantity through PPW-XGboost achieves an MSE of 0.86, which is 0.13 lower than the result without using the PPW loss function.