Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 1
INTRODUCTION
1.1INTRODUCTION
Camouflaged object detection (COD) is one of the most challenging problems in computer
vision, focusing on identifying objects that have evolved or been designed to visually
blend seamlessly into their surrounding environments. Unlike generic object detection
tasks where objects are typically distinguishable from their backgrounds by colour,
texture, or shape, camouflaged objects exploit adaptive colouration, disruptive patterns,
and background mimicry to avoid detection. This phenomenon occurs naturally in the
animal kingdom — for example, stick insects, cuttlefish, and snow leopards — and is also
engineered for military equipment, stealth platforms, and surveillance systems.

The rapid advancements in deep learning and computer vision have opened new
possibilities for automated detection of such concealed objects. However, the intrinsic
difficulty of detecting camouflaged targets means that conventional convolutional neural
network (CNN)-based object detectors frequently fail when confronted with high-
similarity regions between the target and its background. Traditional detection pipelines
relying on contrast cues, edge gradients, or saliency maps are especially ill-suited for
camouflaged scenarios because the very definition of camouflage is the suppression of
such discriminative features.

This project proposes a hybrid transformer-based deep learning framework for detection
of camouflaged objects in complex environments. The framework integrates two
complementary architectures — MilDetr, a military-grade transformer detection module,
and ZoomNeXt, a multi-scale collaborative pyramid network — to achieve accurate
localisation of hidden objects across varying scales, terrains, and occlusion levels. The
system is designed to address the core limitations of existing models: poor generalisation,
high computational cost, and inadequate boundary precision in high-similarity
backgrounds.

Dept. of CSE,SCE 2025-2026 Page 1

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

1.2 Scope of the Project

This project focuses on the design and development of an intelligent deep learning
framework for identifying camouflaged targets in still images using transformer-
based deep learning. The scope encompasses the following key areas:
• Accurate detection of camouflaged objects that visually blend into complex and
heterogeneous backgrounds using a unified hybrid architecture.
• Transformer-based feature extraction to capture fine-grained spatial cues missed by
conventional convolutional detectors.
• Multi-scale feature fusion for robust detection across varying target sizes and
environmental conditions using collaborative pyramid networks.
• A learning strategy that improves sensitivity in high-similarity regions between
targets and backgrounds through uncertainty-aware loss formulation.
• Evaluation across multiple standard benchmark datasets — CAMO, COD10K, and
NC4K — for comprehensive performance comparison.

Dept. of CSE,SCE 2025-2026 Page 2

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 2
LITERATURE SURVEY

1."ZoomNeXt: A Unified Collaborative Pyramid Network for COD", 2024. Youwei Pang et
al.,
This paper proposes ZoomNeXt, a unified collaborative pyramid network designed for
camouflaged object detection. The model employs a “zoom-in/zoom-out” strategy that
integrates multi-scale features with granularity perception units to effectively capture both
global context and fine details. By mimicking human visual perception, the network enhances
the understanding of subtle semantic cues and improves detection of camouflaged objects in
complex backgrounds. The approach demonstrates strong performance across COD
benchmarks, though it may sometimes be influenced by visually salient but irrelevant
background regions.
2."MilDetr: Detection Transformer for Military COD", 2024. Huang Chen,
This paper introduces MilDetr, an end-to-end transformer-based framework tailored for
military camouflaged object detection. The model leverages a Region Selection Network
(RSN) for extracting local features and integrates Feature Pyramid Query (FPQ) mechanisms
to refine object queries. By incorporating local spatial information into the transformer
architecture, the model improves object localization and detection accuracy in challenging
environments. While the approach shows promising results, its evaluation is limited on military
datasets and requires further adaptation for broader natural scene applications.
3. "A ternary encoding network fusing scale awareness and large kernel attention for
camouflaged object detection", Scientific Reports, Nature, 2025. X. Zhuge et al.,
This work introduces a ternary encoding network that fuses scale-awareness and large
kernel attention mechanisms for improved camouflaged object detection. The model
employs multi-scale feature fusion with large receptive fields, achieving superior
performance on standard benchmarks. The study demonstrates that combining scale-
sensitive encoding with large kernel convolutions effectively captures both local texture
and global context, which is critical for detecting objects that closely mimic their
surroundings. The proposed method outperforms prior CNN-based methods on COD10K
and CAMO datasets in terms of both structure measure and boundary precision.

Dept. of CSE,SCE 2025-2026 Page 3

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

4."Uncertainty-Masked Bernoulli Diffusion for Camouflaged Object Detection

Refinement", arXiv preprint, 2025. R. Li et al.,

This paper proposes an uncertainty-masked Bernoulli diffusion model for post-processing

and refinement of camouflaged object detection outputs. The model identifies regions of
high prediction uncertainty and iteratively refines detection masks through a diffusion
process guided by uncertainty maps. This approach is particularly effective for objects
with ill-defined boundaries, which are common in camouflage scenarios. The uncertainty
masking component allows the model to focus refinement on ambiguous boundary regions
while preserving confident predictions in clearly detected regions.

5."Research on Camouflage Target Detection Method Based on Edge Guidance and

Multi-Scale Feature Fusion", Computers, Materials & Continua, Tech Science Press,
2026. W. Zhang et al.,

This study presents a camouflage target detection framework driven by edge guidance
signals combined with multi-scale feature fusion. The edge guidance mechanism explicitly
injects boundary-aware supervision into the training process, enabling the model to learn
precise edge representations even in high-similarity backgrounds. Multi-scale feature
fusion further improves detection robustness across different object sizes. The method is
evaluated on standard camouflaged object benchmarks and demonstrates improved
performance on fine-grained boundary metrics (E-measure and F-measure) compared to
baseline methods.

6."FCL-COD: Weakly Supervised Camouflaged Object Detection with Frequency-

aware and Contrastive Learning", arXiv preprint, 2026. S. Liu et al.,

FCL-COD addresses the challenge of weakly supervised camouflaged object detection by

combining frequency-domain awareness with contrastive learning. The frequency-aware
component analyses the spectral properties of camouflaged regions, which often exhibit
subtle frequency differences from background textures. Contrastive learning encourages
the model to learn discriminative feature representations by pulling together features from
similar camouflaged objects while pushing apart background features. This weakly
supervised approach significantly reduces the need for costly pixel-level annotations while
maintaining competitive detection performance.

7."Diffusion Model for COS with Frequency Domain", 2025. Wei Cai et al.,
This paper presents a diffusion-based model for camouflaged object segmentation (COS)

Dept. of CSE,SCE 2025-2026 Page 4

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

that integrates frequency domain analysis with the denoising diffusion paradigm. The
approach enhances segmentation performance by capturing both spatial and frequency-
based features, allowing the model to better distinguish camouflaged objects from
complex backgrounds. It is particularly effective in scenarios where traditional
discriminative models struggle due to high visual similarity. However, the diffusion
process introduces higher computational complexity, making inference slower compared
to conventional feed-forward networks.

8."SCOUT: Semi-supervised Camouflaged Object Detection by Utilizing Text and

Adaptive Data Selection", IJCAI, 2025. Y. Chen et al.,
SCOUT leverages natural language descriptions in conjunction with visual features
through a semi-supervised learning paradigm for camouflaged object detection. The
system uses text descriptions of camouflage patterns as auxiliary supervision signals,
bridging vision-language understanding with object detection. Adaptive data selection
dynamically prioritises training samples based on difficulty and diversity, improving
model generalisation with limited labelled data. The cross-modal approach introduces a
novel direction for tackling annotation scarcity in camouflaged object detection tasks.

9. "S3OD: Towards Generalizable Salient Object Detection with Synthetic Data",

ICLR, 2026. J. Wang et al.,
S3OD addresses the generalisation problem in salient object detection by leveraging large-
scale synthetic datasets generated through procedural rendering pipelines. The study
shows that models trained on carefully designed synthetic data with realistic lighting,
texture, and occlusion variations can generalise effectively to real-world detection tasks.
The synthetic data generation framework introduces domain randomisation strategies that
expose the model to a wide variety of environmental conditions, improving robustness on
out-of-distribution test sets. These findings directly motivate the synthetic data generation
approach adopted in the present project.

10. "CAMotion: A High-Quality Benchmark for Camouflaged Moving Object Detection

in the Wild", arXiv preprint, 2026. H. Luo et al.,
CAMotion introduces a new large-scale benchmark specifically designed for camouflaged
moving object detection in uncontrolled, real-world environments. The dataset includes
diverse camouflage types, motion patterns, and environmental conditions, providing a
more challenging evaluation setting than existing static image benchmarks. The

Dept. of CSE,SCE 2025-2026 Page 5

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

benchmark reveals significant performance gaps between current state-of-the-art methods

and human-level detection, highlighting the need for models that can integrate temporal
information with spatial camouflage analysis. While the present project focuses on static
images, the benchmark taxonomy provides valuable insights for future extension to video-
based detection.

Dept. of CSE,SCE 2025-2026 Page 6

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 3
EXISTING SYSTEM

Current camouflaged object detection systems can be broadly categorised into CNN-
based methods and transformer-based methods, each with distinct strengths and
limitations.

3.1 CNN-Based Detection Models

CNN-based models such as SINet (Search Identification Network) and PFNet
(Positioning and Focus Network) have established strong baselines for camouflaged
object detection. SINet mimics the predator search strategy, employing a two-stage
pipeline that first searches for coarse object regions and then identifies precise
boundaries. PFNet introduces a positioning and focusing strategy that exploits
attention mechanisms to suppress background distractors. Both models use ResNet or
PVTv2 backbone networks for feature extraction and apply binary cross-entropy
combined with IoU loss during training. While effective on standard benchmarks,
these models struggle with small or highly occluded objects and show poor
generalisation across diverse terrain types with complex textures.

3.2 Transformer-Based Models

Transformer-based models such as CamoFormer and UGTR (Uncertainty-Guided
Transformer) leverage self-attention mechanisms to capture long-range dependencies
in images. CamoFormer employs a hierarchical transformer encoder with window-
based attention to efficiently process high-resolution images. UGTR explicitly
models prediction uncertainty within a transformer framework, providing calibrated
confidence estimates alongside segmentation masks. Video-based approaches further
extend these architectures with temporal attention modules to track camouflaged
objects across frames. However, these models are computationally expensive, require
large GPU memory, and still exhibit accuracy issues near object boundaries due to
insufficient multi-scale spatial sensitivity.

3.3 Limitations of Existing Systems

Dept. of CSE,SCE 2025-2026 Page 7

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

• Systems struggle with small or occluded objects and show poor generalisation across
complex backgrounds.
• Accuracy issues include boundary errors, low-confidence predictions, and missing
fine details near object edges.
• High computational cost and limited real-world annotated data increase the difficulty
of model training and deployment.
• Separate architectures for image and video detection increase system complexity and
maintenance overhead.
• Models often require large training datasets and extensive hyperparameter tuning to
achieve competitive performance.
• Real-time performance is difficult to achieve due to heavy transformer architectures
and high processing overhead.

Dept. of CSE,SCE 2025-2026 Page 8

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 4
PROPOSED SYSTEM

A unified hybrid architecture is proposed for camouflaged object detection that

integrates the complementary strengths of multi-scale pyramid networks and
transformer-based contextual modelling.

4.1 Key Design Principles

• Unified Architecture: A single end-to-end model handles both feature extraction and
detection without separate pre-processing or post-processing modules, reducing system
complexity.
• Collaborative Pyramid Network: Captures mixed-scale semantic information through
zoom-in and zoom-out pathways inspired by the coarse-to-fine nature of human visual
attention.
• Multi-Head Scale Integration Modules: Adaptively aggregate multi-resolution features
using learnable attention weights, balancing global semantic context with local spatial
detail.
• Granularity Perception Units: Apply channel and spatial attention to refine pyramid
features, suppressing background noise and amplifying subtle camouflage-indicative
cues.
• R3FN Transformer Encoder: Strengthens local feature aggregation through region-
level self-attention, capturing dependencies between spatially distributed parts of
concealed objects.
• FPQ Transformer Decoder: Refines learned object queries through cross-attention with
encoder outputs, enabling precise localisation without anchor-based detection.
• Uncertainty-Aware Training: An uncertainty-masked loss function prevents
overconfident predictions in ambiguous boundary regions, improving boundary
precision.

4.2 Advantages of the Proposed System

Dept. of CSE,SCE 2025-2026 Page 9

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

• Improved Detection Accuracy: The hybrid architecture achieves higher accuracy on

camouflaged object benchmarks compared to standalone CNN or transformer models.
• Robust Boundary Precision: Uncertainty-aware loss and granularity perception units
together produce sharper and more accurate object boundary predictions.
• Multi-Scale Robustness: The collaborative pyramid network ensures reliable detection
across small, medium, and large camouflaged objects within the same inference pass.
• Scalable Training: Synthetic data generation through the Blender pipeline supplements
limited real-world annotations, improving model generalisation with less manual
labelling effort.
• Computational Efficiency: Region-level transformer attention is more efficient than
pixel-level attention, reducing computational overhead while maintaining long-range
modelling capability.

Dept. of CSE,SCE 2025-2026 Page 10

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 5
PROBLEM STATEMENT

Camouflaged object detection is difficult due to high visual similarity between the
target and its surrounding background, making it challenging to extract discriminative
features using standard convolutional detectors. The core problems are:

• The absence of distinct foreground-background contrast eliminates reliance on

saliency or edge-based cues that conventional detectors depend upon.

• Existing CNN-based models exhibit poor generalisation across diverse environmental

textures and terrain types, limiting their practical applicability.

• Transformer-based models that address generalisation are computationally expensive

and lack multi-scale spatial sensitivity required for precise boundary detection.

• Limited annotated training data hinders the development of robust supervised learning
models capable of handling the full diversity of camouflage scenarios.

• Current systems treat all pixel predictions equally during training, leading to
overconfident and inaccurate predictions in ambiguous boundary regions.

Therefore, there is a critical need to design and develop a unified hybrid architecture
that integrates multi-scale feature learning, transformer-based contextual modelling,
and uncertainty-aware training strategies to achieve accurate, efficient, and
generalisable camouflaged object detection across diverse real-world scenarios.

Dept. of CSE,SCE 2025-2026 Page 11

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 6
OBJECTIVES

The objectives of this project are as follows:

• To design and develop a unified hybrid architecture for camouflaged object detection
integrating multiscale feature learning and transformer-based contextual modelling.
• To implement a collaborative pyramid network with multi-head scale integration modules
and granularity perception units for robust multi-resolution feature extraction.
• To incorporate transformer-based modules (R3FN and FPQ) for enhanced local feature
aggregation and accurate object localisation in complex backgrounds.
• To evaluate the system based on detection accuracy, boundary precision, and generalisation
performance across diverse camouflage scenarios and benchmark datasets (CAMO,
COD10K, NC4K).
• To address data scarcity through synthetic dataset generation using 3D model-based
rendering pipelines and uncertainty-aware training strategies.

Dept. of CSE,SCE 2025-2026 Page 12

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 7
METHODOLOGY

7.1 Overview
The proposed methodology focuses on developing a unified hybrid deep learning
framework for camouflaged object detection that integrates multi-scale feature
learning with transformer-based contextual modelling. The system combines three
core components: (1) a backbone encoder for initial hierarchical feature extraction,
(2) a Collaborative Pyramid Network (CPN) with zoom-in and zoom-out strategies
for multi-resolution feature aggregation, and (3) transformer-based modules
including the Recurrent Regional Feature Refinement Network (R3FN) and the
Feature Pyramid Query (FPQ) decoder for accurate object localisation in visually
complex backgrounds.

7.2 System Architecture

7.2.1 Architecture Overview
The overall architecture follows an encoder-decoder paradigm. The backbone
network — either a ResNet or Swin Transformer pretrained on ImageNet —
processes the input image to produce hierarchical feature maps at multiple scales
(stride 4, 8, 16, 32). These features capture progressively more abstract semantic
information as depth increases while progressively losing spatial resolution. The
backbone is followed by the Collaborative Pyramid Network, which performs
bidirectional feature aggregation across scales through zoom-in (coarse-to-fine) and
zoom-out (fine-to-coarse) pathways inspired by the human visual system's
hierarchical perception strategy.

7.2.2 Collaborative Pyramid Network (CPN)

The Collaborative Pyramid Network addresses the challenge of detecting

camouflaged objects at varying scales. The zoom-in pathway aggregates high-
resolution local details from shallow feature maps into deeper semantic
representations, preserving fine-grained texture cues that are critical for boundary
delineation. The zoom-out pathway then propagates global semantic context from

Dept. of CSE,SCE 2025-2026 Page 13

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

deep layers back to shallow layers, enriching spatial detail with semantic
understanding.

Multi-Head Scale Integration (MHSI) modules are applied at each pyramid level to
fuse features from adjacent scales using learnable attention weights, allowing the
network to adaptively balance global context and local detail. Granularity Perception
Units (GPU) further refine the aggregated features by applying channel and spatial
attention mechanisms that suppress background noise and amplify subtle target-
indicative cues.

Fig 1.1 Architecture of ZoomNeXt

7.2.3 Transformer-Based Contextual Modelling

The Recurrent Regional Feature Refinement Network (R3FN) operates as a

transformer encoder that processes multi-scale feature queries through stacked self-
attention and feed-forward layers. Unlike standard ViT-style transformers that
operate on fixed patch tokens, R3FN employs region-level attention computed over
semantically coherent spatial patches derived from the CPN output.

The Feature Pyramid Query (FPQ) module serves as the transformer decoder, taking
the R3FN output as key-value pairs and iteratively refining a set of learned object
queries through cross-attention. The refined queries are then decoded into
segmentation masks and bounding box predictions without relying on anchor boxes
or non-maximum suppression.

Dept. of CSE,SCE 2025-2026 Page 14

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

Fig 1.2 Architecture of MilDetr

7.3 Dataset Preparation

7.3.1 Benchmark Datasets
Three standard benchmark datasets are employed for training and evaluation:
• CAMO: Contains 1,250 camouflaged images and 1,250 non-camouflaged images
across 8 categories, annotated with pixel-level ground truth masks.
• COD10K: The largest COD benchmark with 10,000 images spanning 78 camouflage
sub-categories, with hierarchical annotations including body, edge, and instance-level
masks.
• NC4K: A test-only dataset of 4,121 images providing a challenging out-of-
distribution evaluation setting for assessing model generalisation.

7.3.2 Synthetic Dataset Generation

To address annotated data scarcity, a synthetic data generation pipeline is developed
using Blender. High-fidelity 3D models of camouflaged objects are textured with
procedurally generated camouflage patterns and composited onto realistic
background scenes with controlled lighting, shadows, and atmospheric effects.
Domain randomisation strategies vary object scale, orientation, illumination, and
background texture, producing a diverse synthetic training set.

7.3.3 Uncertainty-Aware Loss Function

The proposed uncertainty-aware loss function incorporates a learned uncertainty map
that assigns lower training weights to boundary pixels where the model prediction is

Dept. of CSE,SCE 2025-2026 Page 15

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

inherently uncertain. This is implemented as a custom PyTorch loss module that

jointly optimises the detection mask and the uncertainty map, preventing the model
from being penalised excessively for ambiguous annotations near object boundaries.

7.4 Training Process

The model is trained end-to-end using the AdamW optimiser with a weight decay of
1e-4 and an initial learning rate of 1e-4, following a cosine annealing schedule with
warm restarts. The backbone is initialised with pretrained ImageNet weights. Data
augmentation strategies include random horizontal flipping, random cropping, colour
jitter, and mixup augmentation with background images. The batch size is set to 8
with gradient accumulation over 4 steps.

7.5 System Design Flowchart

The complete system pipeline: (1) Input image fed to backbone encoder for
hierarchical feature extraction. (2) Multi-scale feature maps passed through CPN for
zoom-in and zoom-out aggregation. (3) MHSI modules and Granularity Perception
Units refine pyramid features. (4) R3FN encoder applies transformer self-attention
over regional feature tokens. (5) FPQ decoder iteratively refines object queries via
cross-attention with encoder outputs. (6) Final detection masks and bounding box
predictions decoded from refined queries. (7) Post-processing applies uncertainty-
based mask refinement to sharpen boundary predictions. (8) Evaluation metrics
computed against ground truth annotations.

Fig 1.3 Architecture

Dept. of CSE,SCE 2025-2026 Page 16
Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 8
SYSTEM REQUIREMENTS

8.1 Hardware Requirements

Component Specification
Processor Intel Core i7 / AMD Ryzen 9 (10th Gen or above)
RAM 16 GB minimum (32 GB recommended)
GPU NVIDIA RTX 3080 / A100 (CUDA 11.7+, 16 GB
VRAM recommended)
Storage 512 GB SSD (NVMe recommended for fast
dataset I/O)
Operating System Ubuntu 20.04 LTS / Windows 11 with WSL2

8.2 Software Requirements

8.2.1 Programming and Core Frameworks
• Python 3.10+ — End-to-end implementation of the detection pipeline, data
preprocessing, training, and evaluation scripts.
• PyTorch 2.0+ — Core deep learning framework for hybrid architecture training,
transformer module implementation, and GPU-accelerated computation.

8.2.2 Detection Architecture and Modules

• timm (PyTorch Image Models) — Backbone network support including ResNet
variants and Swin Transformer for pretrained feature extraction.
• torchvision — Data loading utilities, standard augmentation pipelines, and image
preprocessing functions.
• MMCV / MMDetection (optional) — Modular detection framework for rapid
prototyping of detection head configurations.

8.2.3 Dataset and Annotation

• CAMO, COD10K, NC4K — Standard benchmark datasets for camouflaged object
detection with pixel-level ground truth annotations.
• Blender (v3.5+) / Synthetic pipeline — 3D model-based synthetic dataset generation
for data augmentation and domain randomisation.

Dept. of CSE,SCE 2025-2026 Page 17

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

8.2.4 Loss and Uncertainty Modelling

• Custom PyTorch loss module — Implements the uncertainty-aware loss function with
jointly optimised detection mask and uncertainty map outputs.
• scipy — Statistical functions supporting uncertainty quantification and evaluation
metric computation.

8.2.5 Visualisation and Analysis

• Matplotlib and Seaborn — Performance metric plotting, training curve visualisation,
and qualitative result comparison figures.
• OpenCV — Prediction mask visualisation, contour drawing for boundary analysis, and
qualitative detection result generation.

8.2.6 Experiment Tracking and Utilities

• TensorBoard — Real-time training monitoring including loss curves, learning rate
schedules, and detection visualisation at checkpoints.
• Pandas — Evaluation data management, metric aggregation across datasets, and
ablation study result organisation.
• tqdm — Training progress tracking with per-iteration speed and estimated time-to-
completion reporting.
• Weights and Biases (optional) — Cloud-based experiment tracking, hyperparameter
sweeping, and run comparison.

Dept. of CSE,SCE 2025-2026 Page 18

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

CHAPTER 9
CONCLUSION

9.1 Summary of Phase-I Progress

This Phase-I report presents the design and initial development of a unified hybrid
deep learning framework for detection of camouflaged objects in complex
environments. The work completed in Phase-I encompasses a thorough literature
survey of state-of-the-art camouflaged object detection methods, a detailed analysis
of the limitations of existing CNN-based and transformer-based approaches, and the
complete architectural design of the proposed system.

The proposed architecture integrates a Collaborative Pyramid Network (CPN) with

Multi-Head Scale Integration modules and Granularity Perception Units for multi-
resolution feature aggregation, coupled with transformer-based modules (R3FN
encoder and FPQ decoder) for contextual modelling and precise object localisation.
An uncertainty-aware loss function has been designed to handle the inherent
ambiguity in annotating camouflaged object boundaries. The system requirements,
dataset strategy (including synthetic data generation via Blender), and evaluation
framework using CAMO, COD10K, and NC4K benchmarks have been fully
specified.

9.2 Evaluation Metrics

The proposed system will be evaluated using four standard camouflaged object
detection metrics:
• Structure Measure (Sα): Evaluates structural similarity between predicted segmentation
mask and ground truth. A higher Sα value (range 0–1) indicates better structural
correspondence.
• Enhanced Alignment Measure (Eφ): Evaluates both local pixel-level accuracy and
global image-level statistics simultaneously, providing a holistic measure sensitive to
boundary precision.
• Mean Absolute Error (MAE): Computes per-pixel absolute difference between
predicted probability map and binary ground truth mask. Lower MAE indicates more
accurate prediction.

Dept. of CSE,SCE 2025-2026 Page 19

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

• F-measure (Fβ): A weighted harmonic mean of precision and recall, computed with β²
= 0.3 to emphasise precision, balancing false positives and false negatives.

9.3 Future Implementation Plan

The following tasks are planned for Phase-II implementation:

• Complete implementation of the full hybrid architecture in PyTorch, including the
CPN, R3FN encoder, and FPQ decoder modules.
• Develop the synthetic data generation pipeline in Blender with domain
randomisation for camouflage pattern diversity.
• Train the model on CAMO and COD10K training splits using the uncertainty-aware
loss function and AdamW optimiser.
• Conduct comprehensive evaluation on CAMO, COD10K, and NC4K test splits,
measuring Sα, Eφ, MAE, and Fβ metrics.
• Perform ablation studies to validate the individual contribution of each architectural
module (CPN, MHSI, GPU, R3FN, FPQ).
• Compare results against state-of-the-art COD models including SINet-V2, PFNet,
CamoFormer, and UGTR.
• Optimise model for computational efficiency and explore lighter backbone
alternatives for resource-constrained deployment.

Dept. of CSE,SCE 2025-2026 Page 20

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

REFERENCES

[1] "A ternary encoding network fusing scale awareness and large kernel attention for
camouflaged object detection," Scientific Reports, Nature, 2025.
[2] "Uncertainty-Masked Bernoulli Diffusion for Camouflaged Object Detection Refinement,"
arXiv preprint, 2025.
[3] "Research on Camouflage Target Detection Method Based on Edge Guidance and Multi-
Scale Feature Fusion," Computers, Materials & Continua, Tech Science Press, 2026.
[4] "FCL-COD: Weakly Supervised Camouflaged Object Detection with Frequency-aware and
Contrastive Learning," arXiv preprint, 2026.
[5] "SCOUT: Semi-supervised Camouflaged Object Detection by Utilizing Text and Adaptive
Data Selection," IJCAI, 2025.
[6] "S3OD: Towards Generalizable Salient Object Detection with Synthetic Data," ICLR, 2026.
[7] "CAMotion: A High-Quality Benchmark for Camouflaged Moving Object Detection in the
Wild," arXiv preprint, 2026.
[8] "Uncertainty-Aware Transformer for Referring Camouflaged Object Detection," IEEE
Transactions on Image Processing, 2025.
[9] Y. Fan et al., "Concealed Object Detection," IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 44, no. 10, pp. 6024–6042, 2022.
[10] T. Mei et al., "Camouflaged Object Detection with Feature Decomposition and Edge
Reconstruction," IEEE CVPR, 2023.
[11] J. Wang et al., "PFNet: Progressive Feature Fusion for Camouflaged Object Detection,"
IEEE Transactions on Image Processing, 2021.
[12] H. Lv et al., "Simultaneously Localize, Segment and Rank the Camouflaged Objects," IEEE
CVPR, 2021.
[13] Z. Liu et al., "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,"
ICCV, 2021.
[14] N. Carion et al., "End-to-End Object Detection with Transformers (DETR)," ECCV, 2020.

Dept. of CSE,SCE 2025-2026 Page 21

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

PUBLISHED LITERATURE SURVEY PAPER

Dept. of CSE,SCE 2025-2026 Page 22

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

Dept. of CSE,SCE 2025-2026 Page 23

Deep Learning Framework for Detection of Camouflaged Objects in Complex Environments

Dept. of CSE,SCE 2025-2026 Page 24