ResNet-Integrated Gradient Guided Multi-Scale Network for Camouflaged Object Detection

Authors

  • Jogeswar Behera Post Graduate (MCA), Radhakrishna Institute of Technology and Engineering, Bhubaneswar, Odisha, India; affiliated to Biju Patnaik University of Technology, Rourkela, Odisha, India. Author
  • Kushal Minz Post Graduate (MCA), Radhakrishna Institute of Technology and Engineering, Bhubaneswar, Odisha, India; affiliated to Biju Patnaik University of Technology, Rourkela, Odisha, India. Author
  • Dibyajyoti Sethi Post Graduate (MCA), Radhakrishna Institute of Technology and Engineering, Bhubaneswar, Odisha, India; affiliated to Biju Patnaik University of Technology, Rourkela, Odisha, India. Author
  • Priyanka Shit Assistant Professor, Radhakrishna Institute of Technology and Engineering, Bhubaneswar, Odisha, India; Biju Patnaik University of Technology, Rourkela, Odisha, India. Author
  • Jashasmita Pal Assistant Professor, Radhakrishna Institute of Technology and Engineering, Bhubaneswar, Odisha, India; Biju Patnaik University of Technology, Rourkela, Odisha, India. Author

DOI:

https://doi.org/10.47392/IRJASH.2026.040

Keywords:

Camouflaged Object Detection (COD), Computer Vision, Deep Learning, Feature Fusion, Gradient Guidance

Abstract

Camouflaged Object Detection (COD) is a challenging computer vision task that aims to identify objects that visually blend into their surrounding environment. Traditional object detection techniques often struggle with camouflage because the foreground and background may share similar color, texture, shape, and local contrast. This paper presents a ResNet-Integrated Gradient Guided Multi-Scale Network (RIGGMNet) that combines deep residual feature extraction, gradient-guided boundary enhancement, and multi-scale feature fusion. ResNet is used as the backbone to obtain progressively richer semantic representations, while a gradient-guided module emphasizes subtle boundaries and texture transitions. Multi-scale aggregation then combines fine spatial details with high-level contextual information so that objects of different sizes and complexities can be segmented more reliably. The network is designed as an end-to-end segmentation framework that produces a dense binary map of the camouflaged region. Architecturally, the framework is compared against representative COD baselines to situate its design choices within the broader literature, and it is positioned as a practical approach for wildlife monitoring, surveillance, medical imaging, autonomous systems, and environmental research. Consistent with the source research abstract, exact numerical benchmark scores are left to a reproducible experimental run rather than being invented for this report.

Downloads

Published

2026-11-21