Camouflaged Object Detection. Masters thesis, King Fahd University of Petroleum and Minerals.

PDF (Masters Thesis) BaberJan_202213840_COD_Thesis (1).pdf - Submitted Version Restricted to Repository staff only until 13 July 2026. Download (99MB)

Arabic Abstract

يرهوجلا هباشتلا :ةيساسٔا تايدحت ةتس هجاويو ،اهطيحم عم ةسالسب جزتمت يتلا ماسجٔالا ديدحت ىلٕا (COD) ةهومملا ماسجٔالا فاشتكا فدهي ماسجٔالا زييمتو (CD)، ةيقايسلا تايدامتعالاو (EC)، ةيئيبلا تاديقعتلاو (ESV)، مجحلا يف ةديدشلا تانيابتلاو (ED)، فاوحلا بارطضاو (IS)، قرطلا يناعت ،يعانصلا صحفلاو ةيربلا ةايحلا ةبقارمو يبطلا ريوصتلا يف ةيويحلا تاقيبطتلل ةردقلا هذه ةيمهٔا مغر (SCOD). ةزرابلا نع ةهومملا SCOD تايلٓا بايغو ،لماش لكشب ةتسلا تايدحتلا ةجلاعم ىلع ةردقلا مدعو ،يرزٓات لماكت نود يرامعملا مكارتلا :ةيجهنم دويق ثالث نم ةيلاحلا .يجيردتلا يرامعملا ريوطتلا لالخ نم يرزٓاتلا ميمصتلا ئدابم ثحبلا اذه تبث ُ ي ،تاوجفلا هذه ةجلاعمل .ةيجراخ تادامتعا بلطتي امم ةيلخادلا لالخ نم EDو IS ةيساسٔالا تايدحتلا لحل قسانتب لمعت (PEDو EFEو (CFI ةلماكتم تادحو ثالث نٔا رهظُ ي يذلا SPEGNet ً الؤا انروط (ةيناث/راطٕا +60) يلعفلا تقولا يف ةجلاعم عم COD10K ىلع 890 .0=Sα ً اققحم, يشالتلاو ةورذلا طمنب فاوحلا ليدعت هيجوتلا ةيلٓا ً امدقم ،راسملا ةيئانث ةجلاعم ربع ةلماش ةيطغتل ةيرزٓاتلا ئدابملا C3Net عسوي ،ققحتلا اذه ىلع ً ءانب .طقف لماعم نويلم 29 .3 مادختساب COD10K، ىلع 896 .0=Sα) ً اروطتم ً ءادٔا ً اققحم ،ةيجراخ تادامتعا نود SCOD عم يلخادلا لماعتلل (ICG) يرصبلا قايسلا ىلع دمتعملا ةرورض تبثي امم ،(7% .35 ةوجفب ،549 .0=Sα (SAM2: ةيساسٔالا جذامنلا ىلع ريبك لكشب نيرامعملا الك قوفتي NC4K). ىلع 910 .0 يلعفلا تقولا يف ةردقلا ىلع ظافحلاب روطتملا ءادٔالا يف SPEGNet مدقت ،يجيردتلا روطتلا اذه لالخ نم COD. ـل ةصصختملا تايرامعملا SCOD ةردق كلذ يف امب ةتسلا COD تايدحت عيمج عم لماشلا لماعتلا يف لؤالا هنوكب رثكٔا دودحلا C3Net عفد امنيب ،ةيساسٔالا تايدحتلل.ةزرابلا تاتتشملا عم ةدقعملا تاهويرانيسلا يف يلمعلا رشنلا حيتي امم ،ةيلخادلا

English Abstract

Camouflaged Object Detection (COD) aims to identify objects that blend seamlessly with their surroundings, facing six fundamental challenges: Intrinsic Similarity (IS), Edge Disruption (ED), Extreme Scale Variations (ESV), Environmental Complexities (EC), Contextual Dependencies (CD), and Salient-Camouflaged Object Disambiguation (SCOD). While this detection capability enables critical applications in medical imaging, wildlife monitoring, and industrial inspection, current methods suffer from three systematic limitations: architectural accumulation without synergistic integration, inability to comprehensively address all six challenges, and lack of intrinsic SCOD mechanisms requiring external dependencies. To address these gaps, this research validates synergistic design principles through progressive architectural development. We first developed SPEGNet, which demonstrates that three complementary modules (CFI, EFE, PED) working in concert can solve core challenges IS and ED through peak-and-fade edge modulation (20%→33%→0%), achieving Sα=0.890 on COD10K with real-time inference (60+ FPS) using only 3.29M parameter contributions. Building on this validation, C3Net extends synergistic principles to comprehensive coverage through dual-pathway processing, introducing the Image-based Context Guidance (ICG) mechanism for intrinsic SCOD handling without external dependencies, attaining state-of-the-art performance (Sα=0.896 on COD10K, 0.910 on NC4K). Both architectures significantly outperform foundation models (SAM2: Sα=0.549, representing a 35.7% gap), proving that specialized architectures are necessary for COD. Through this progressive development, SPEGNet advanced state-of-the-art by maintaining real-time capability for core challenges, while C3Net pushed boundaries further by becoming the first to comprehensively handle all six COD challenges including intrinsic SCOD capability, enabling practical deployment in complex scenarios with salient distractors.

Item Type:	Thesis (Masters)
Subjects:	Computer Seminars Research Research > Information Technology Research > Engineering
Department:	College of Computing and Mathematics > Computer Engineering
Committee Advisor:	El-Maleh, Aiman H.
Committee Members:	Anwar, Saeed and Siddiqui, Abdul Jabbar
Depositing User:	BABER JAN (g202213840)
Date Deposited:	13 Jul 2025 06:21
Last Modified:	13 Jul 2025 06:21
URI:	http://eprints.kfupm.edu.sa/id/eprint/143612