Article: Morphology extraction from blurred image targets via deep multi-class modelling Journal: International Journal of Information and Communication Technology (IJICT) 2025 Vol.26 No.33 pp.91 - 106 Abstract: Motion blur and defocus blur significantly distort the morphology of image targets. This paper proposes multi-task fusion network (MTF-Net), a MTF-Net based on deep learning, to address this challenge. The network incorporates an adaptive blur awareness module (ABAM) to characterise degradation patterns. A parameter-shared dual-path architecture enables simultaneous target classification and morphological segmentation, achieving end-to-end category recognition and precise contour extraction. Comprehensive evaluations on BSD-Blur, RealBlur-R, and COCO-Blur datasets demonstrate the model's effectiveness: it attains 94.3% classification accuracy (surpassing Deeplabv3+ by 7.2%), elevates the morphological similarity index (MSI) by 12.7%, and processes images at 23 FPS. Ablation studies confirm ABAM's critical role in blur robustness (removal causes an 8.2% MSI performance drop). MTF-Net provides a high-precision solution for applications demanding accurate shape perception, including medical imaging and autonomous driving. Inderscience Publishers - linking academia, business and industry through research

Title: Morphology extraction from blurred image targets via deep multi-class modelling

Authors: Shuo Feng

Addresses: College of Information and Intelligence, Hunan Agricultural University, Changsha 410128, China

Abstract: Motion blur and defocus blur significantly distort the morphology of image targets. This paper proposes multi-task fusion network (MTF-Net), a MTF-Net based on deep learning, to address this challenge. The network incorporates an adaptive blur awareness module (ABAM) to characterise degradation patterns. A parameter-shared dual-path architecture enables simultaneous target classification and morphological segmentation, achieving end-to-end category recognition and precise contour extraction. Comprehensive evaluations on BSD-Blur, RealBlur-R, and COCO-Blur datasets demonstrate the model's effectiveness: it attains 94.3% classification accuracy (surpassing Deeplabv3+ by 7.2%), elevates the morphological similarity index (MSI) by 12.7%, and processes images at 23 FPS. Ablation studies confirm ABAM's critical role in blur robustness (removal causes an 8.2% MSI performance drop). MTF-Net provides a high-precision solution for applications demanding accurate shape perception, including medical imaging and autonomous driving.

Keywords: blurred image; target shape extraction; multi-classification model; adaptive perception module; morphological similarity index; MSI.

DOI: 10.1504/IJICT.2025.148654

International Journal of Information and Communication Technology, 2025 Vol.26 No.33, pp.91 - 106

Received: 30 Jun 2025
Accepted: 23 Jul 2025
Published online: 17 Sep 2025 *

Title: Morphology extraction from blurred image targets via deep multi-class modelling

Keep up-to-date