Segmentation of flexible intermediate bulk container lifting lug for electrolytic aluminum plant based on improved U-Net

This paper addresses the automation of flexible intermediate bulk container (FIBC) lifting loop segmentation in aluminum electrolysis plants by proposing an improved semantic segmentation approach based on an enhanced U-Net architecture. Traditional manual operations are inefficient and pose safety risks, while factors such as dust and uneven lighting in complex industrial environments challenge existing vision algorithms. To overcome these issues, this study introduces three attention mechanisms—SimAM, ECA, and PSA—to strengthen the model’s ability to extract edge and texture features of the lifting loops. Additionally, a Haar wavelet-based downsampling module replaces conventional pooling to preserve high-frequency detail information. Experimental results demonstrate that the U-Net-ECA-Haar model achieves mean Intersection over Union (mIoU), mean Pixel Accuracy (mPA), and mean Precision (mPrecision) of 89.23%, 93.32%, and 94.69% respectively on the validation set, which are 3.06%, 2.36% and 1.54% higher than the original U-Net model, respectively. Furthermore, the creation of the first FIBC lifting loop dataset provides a foundational resource for future research. This study offers an efficient solution for target segmentation tasks in industrial scenarios, with substantial practical application value.