Skip to content
Open access

A DeepLabV3-Based Network With Hybrid Loss and Adaptive Inference for Camouflaged Object Detection

2026 · IEEE Access · Vol 14, pp. 166110-166138 · 0 citations · 45 references

TL;DR

DeepCamNet, a DeepLabV3-based framework that examines representation capacity, composite supervision, and inference efficiency through two independently trained configurations: ResNet101 and MobileNetV3, is presented.

Abstract

Camouflaged object detection (COD) requires segmenting objects with weak foreground–background contrast and ambiguous boundaries. This paper presents DeepCamNet, a DeepLabV3-based framework that examines representation capacity, composite supervision, and inference efficiency through two independently trained configurations: ResNet101 and MobileNetV3. Atrous spatial pyramid pooling aggregates contextual information, while a hybrid Dice–binary-cross-entropy–focal objective combines region overlap, pixel-wise discrimination, and hard-pixel emphasis. Paired spatial augmentation maintains image–mask alignment. A decision threshold is selected on validation data and fixed during testing; the inference rule is therefore not image-adaptive. The study reports static-image comparisons on CAMO, CHAMELEON, COD10K, and NC4K, together with frame-wise CAD and MoCA evaluations, convergence plots, qualitative examples, and scale/pooling comparisons. The reported ResNet101 configuration obtains <inline-formula> <tex-math notation="LaTeX">$S_{m}=0.9241$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$F_{\beta }^{w}=0.9185$ </tex-math></inline-formula>, and <inline-formula> <tex-math notation="LaTeX">$E_{m}=0.9640$ </tex-math></inline-formula> on CAMO; <inline-formula> <tex-math notation="LaTeX">$S_{m}=0.9480$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$E_{m}=0.9960$ </tex-math></inline-formula> on CHAMELEON; <inline-formula> <tex-math notation="LaTeX">$S_{m}=0.9390$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$F_{\beta }^{w}=0.9360$ </tex-math></inline-formula>, and <inline-formula> <tex-math notation="LaTeX">$\mathrm {MAE}=0.0150$ </tex-math></inline-formula> on COD10K; and <inline-formula> <tex-math notation="LaTeX">$S_{m}=0.9270$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$E_{m}=0.9820$ </tex-math></inline-formula> on NC4K. MobileNetV3 provides a complementary configuration with competitive reported scores and higher reported throughput. These comparisons motivate an application-oriented accuracy–efficiency study rather than a claim of a new backbone or loss family. Their interpretation is subject to the cross-table metric and configuration-consistency limitations identified in the experimental analysis.

Read PDF

Similar papers

Preprint Oct 2026

AxisQ: Depth-Adaptive Channel--Frequency Quantization for Camouflaged Object Detection

Camouflaged object detection (COD) depends on weak texture and boundary cues that can be destroyed by aggressive post-training quantization (PTQ). Existing PTQ methods mainly optimize quantizers or reconstruct network blocks, while using a single precision-allocation domain throughout hierarchical backbones. Instead, w...

Chiung-Wen Ruan, Tian-Qin Li, Xin He et al. · 0 citations
Open access Oct 2026

Camouflaged Object Segmentation Using a U-Net Based Deep Learning Approach

The paper focuses on the computer vision task of camouflaged object detection that has various applications, including, but not limited to, military, security, biology, and industry. From the perspective of machine vision, camouflage is a state in which the target’s texture, colour, and pattern are indistinguishable fr...

D. Devi, Arpita Prusty, Deepanjali Khadia et al. · 0 citations
Open access Oct 2026

Deepfake Image Detection Using Squeeze-and-Excitation Attention and Adaptive Threshold Optimization

Artificial intelligence-generated content (AIGC) has significantly improved the realism of manipulated facial images and videos, creating serious risks for social-network misinformation, content security, and digital trust. This study focuses on visual deepfake image detection rather than multimodal misinformation dete...

Chao-Ran Li · 0 citations
Book Open access Nov 2026

CT-CLIP: Calibrating Semantic Occupancy Tokens for Class-Guided Camouflaged Object Detection

CT-CLIP is proposed, a novel framework for class-guided camouflaged object detection that calibrates SOT to restore spatial structure and significantly outperforms existing methods on the CGCOD benchmarks.

Guo-Hui Zhang, Fu-Ming Sun, Xing-Rui Zhou et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.