DDR-Unet: A High-Accuracy and Efficient Ore Image Segmentation Method
Fei Li, Xiaoyan Liu, Yufeng Yin, Zongping Li
Abstract
This paper presents DDR-Unet, a high-accuracy and efficient method for ore image segmentation (OIS). OIS is a crucial step in measuring ore particle size distribution (PSD), but it faces challenges due to variations in particle sizes, shapes, overlaps, and powder interference. DDR-Unet improves U-Net in three aspects: encoder, decoder, and loss function. The encoder adopts deformable convolution to capture ore features of different sizes and shapes. The decoder employs multi-level dense residual connections to fuse low-level and high-level features. The loss function uses weight-adaptive BCE to balance the number of ore and non-ore samples. We evaluate DDR-Unet on an ore dataset and two public datasets and compare it with ten state-of-the-art segmentation methods. DDR-Unet outperforms all methods in OIS performance and PSD error. The code is available at: https://github.com/lifeiwen/DDR-Unet-for-ore-image-segmentat.