MHAED-Net: a lightweight multiscale hybrid attention encoder-decoder network for the efficient segmentation of industrial forging images

Accurate and efficient segmentation of the boundaries, shapes, and sizes of forgings is crucial for intelligent forging perception. Current image segmentation techniques frequently face challenges in achieving an effective balance between speed and accuracy. Moreover, these techniques often fail to...

Full description

Saved in:
Bibliographic Details
Published in:The Journal of supercomputing Vol. 81; no. 8; p. 1001
Main Authors: Wan, Miao, Lin, Y. C., Li, Shu-Xin, Wu, Gui-Cheng, Zeng, Ning-Fu, Zhang, Song, Chen, Ming-Song, Li, Chao, Zhan, Xiao-Dong, Qiu, Yu-Liang
Format: Journal Article
Language:English
Published: New York Springer US 09.06.2025
Springer Nature B.V
Subjects:
Accuracy
Compilers
Computer Science
Datasets
Decision making
Deep learning
Efficiency
Encoders-Decoders
Energy consumption
Floating point arithmetic
Forging
Forgings
Image segmentation
Industrial applications
Interpreters
Machine learning
Manufacturing
Methods
Neural networks
Processor Architectures
Programming Languages
Working conditions
Forging process
Intelligent manufacturing
Image segmentation
Visual monitoring
Lightweight network
ISSN:1573-0484, 0920-8542, 1573-0484
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate and efficient segmentation of the boundaries, shapes, and sizes of forgings is crucial for intelligent forging perception. Current image segmentation techniques frequently face challenges in achieving an effective balance between speed and accuracy. Moreover, these techniques often fail to adapt well to the complex working conditions and diverse scales of forgings. In this study, a lightweight multiscale hybrid attention encoder-decoder network (MHAED-Net) is designed for the efficient segmentation of industrial forging images. MHAED-Net is characterized by only 0.076 M parameters and 0.087 Giga Floating-point Operations Per Second. The model employs a novel multiscale hybrid attention block (MHAB) that integrates the convolution normalization activation block and the ShuffleNetV2 block to create an encoder-decoder network. The proposed MHAB integrates CNN-based and Transformer-based attention through a multi-branch fusion approach. It employs dilated convolutions for multi-scale feature learning and incorporates a Lightweight Transformer to capture long-range dependencies. MHAED-Net achieves a mean Intersection over Union of 94.82% and Dice Similarity Coefficient of 97.34% in the segmenting the FORSeg dataset. Extensive experimental results demonstrate that MHAED-Net achieves state-of-the-art performance under complex conditions and multi-scale scenarios, highlighting its significant potential for industrial applications.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1573-0484
0920-8542
1573-0484
DOI:10.1007/s11227-025-07456-8