Experimentation Prompts for Anomaly Detection (AD)

So that reviewers and readers can reproduce the results for this IEEE paper, please find the various "Experimentation Prompts for Anomaly Detection (AD)" below (rather than in the paper). They are numbered AD #1 through AD #16.

AD #1: Stage-Based Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Stage-based reasoning, wherein each stage serves as a discrete step so that the involved AI model can plan prior to the next step. It should be noted that in markdown, **text** signifies bold; theoretically, the involved AI model should weight the associated instructions slightly higher.

Task:

Perform a detailed analysis of the uploaded image to produce a simplified, annotated version.

 

Stage 1: Object Detection

- Detect all major objects and structural elements.

- Categorize each object as: primary structure, background element, vehicle, human, or other.

- Flag any objects or patterns that appear unusual or anomalous relative to the scene.

 

Stage 2: Image Simplification

- Generate a simplified version of the image that preserves:

- essential shapes

- layout

- primary objects

- Remove minor details like textures, small artifacts, and noise.

 

Stage 3: Overlay Application

- Overlay bright outlines on all detected objects.

- Use **red outlines specifically for anomalies**.

- Ensure overlays are clearly visible and do not obscure the rest of the image.

 

Stage 4: Feature Description

- Provide a short description for each outlined object including:

- Object category

- Position/location in the image

- Reason it was detected

- Any anomaly status

 

Stage 5: Output Formatting

- Return the modified image with overlays applied.

- Provide a concise table or bullet list of all features and anomalies.

- Maintain the original image layout and spatial relationships.

- Ensure descriptions are 1–2 sentences per object.

 

AD #2: Heuristical Facilitation Prompt

(as referenced in Section IIIC. Experimentation)

Value-added Proposition: The specification for bright overlays may facilitate heuristical detection.

Task: Analyze the uploaded image and produce a simplified, annotated version.

 

Steps:

1. Detect all major objects and structural elements. Categorize each as: primary structure, background, vehicle, human, or other. Flag unusual/anomalous objects.

 

2. Simplify the image by preserving only essential shapes, layout, and primary objects. Remove minor details and noise.

 

3. Overlay bright outlines around every detected object. Use red outlines for anomalies. Keep the rest of the image unchanged.

4. Provide a concise list of all features with:

- Object category

- Approximate location

- Reason for detection

- Anomaly status (if any)

 

Constraints:

- Preserve original spatial layout.

- Do not alter the base image except for overlays.

- Keep descriptions 1–2 sentences per object.

- Ensure overlays are visually clear and distinguishable.

 

AD #3: Major Features and Candidates Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: The specification of a prospective candidate list of major features may facilitate heuristical matters.

Task:

Analyze the uploaded satellite night-city image and produce a simplified, annotated version.

 

Steps:

1. Detect all major urban features, including:

- Buildings / downtown clusters

- Airports and airstrips

- Major highways and interchanges

- Bridges, rivers, and waterways

- Parks or open spaces

 

2. Categorize each detected object: primary structure, transportation, natural feature, or other.

 

3. Identify anomalies, such as unusual shapes, missing patterns in grids, or irregular lighting; highlight these in red.

 

4. Simplify the image by keeping only essential shapes, layout, and primary objects. Remove minor details, textures, or noise.

 

5. Overlay bright, high-contrast outlines on every detected object; use **red outlines specifically for anomalies**.

 

6. Provide a concise list describing each outlined feature:

- Object category

- Approximate location

- Reason for detection

- Anomaly status if applicable.

 

Constraints:

- Preserve original spatial layout and scale.

- Do not alter the base image except for overlays.

- Keep descriptions brief (1–2 sentences per feature).

- Ensure overlays are visually distinct and easy to interpret.

 

AD #4: Careful Reasoning Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Obligates the involved AI model to carefully analyze, reason, and be as transparent as possible, particularly in the area of false positives and false negatives via multiple reasoning paths and cross-checking. This prompt also undertakes rearticulation via reiteration with "highlight anomalies with **red outlines**" and "Red outlines are reserved for anomalies."

Task:

Analyze the uploaded satellite night-city image and produce a simplified, annotated version.

 

Instructions:

- Think carefully and reason step-by-step before producing your answer.

- Generate multiple reasoning paths, cross-check for consistency, and note assumptions or uncertainties.

- Analyze potential false positives (objects incorrectly identified) and false negatives (missed objects), define each clearly, and assess their likelihood and consequences in the context of urban and power infrastructure features.

- Compare trade-offs of including vs. excluding objects or anomalies.

- Provide examples where relevant to clarify reasoning.

 

Steps:

1. Object Detection:

- Detect all major urban features, including:

- Buildings / downtown clusters

- Airports and airstrips

- Major highways and interchanges

- Bridges, rivers, and waterways

- Parks or open spaces

- Categorize each detected object: primary structure, transportation, natural feature, or other.

 

2. Anomaly Detection:

- Identify anomalies, such as unusual shapes, missing grid patterns, irregular lighting, or inconsistencies in infrastructure.

- Highlight anomalies with **red outlines**.

 

3. Image Simplification:

- Keep only essential shapes, layout, and primary objects.

- Remove minor details, textures, and noise.

 

4. Overlay Instructions:

- Overlay bright, high-contrast outlines on all detected objects.

- Ensure overlays are visually distinct and do not obscure the base image.

- Red outlines are reserved for anomalies.

 

5. Feature Description:

- Provide a concise list describing each outlined feature:

- Object category

- Approximate location

- Reason for detection

- Anomaly status if applicable

- Include discussion of potential false positives and false negatives with examples.

 

6. Summary and Reasoning:

- Summarize findings based on the most consistent reasoning across multiple paths.

- Note assumptions, uncertainties, likelihoods, and potential consequences.

- Provide a short conclusion highlighting the most reliable insights.

 

Constraints:

- Preserve original spatial layout and scale.

- Do not alter the base image except for overlays.

- Keep descriptions brief (1–2 sentences per object).

- Ensure overlays and highlights are easy to interpret.

 

AD #5: Quality Assurance/Quality Control (QA/QC) Feature Mapping Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Facilitates QA/QC by obligating the involved AI model to conduct feature mapping, thereby further mitigating against false positives and/or false negatives.

Task:

Analyze the uploaded satellite night-city image and produce a simplified, annotated version.

 

Instructions:

- Think carefully and reason step-by-step before producing your answer.

- Generate multiple reasoning paths, cross-check for consistency, and note assumptions or uncertainties.

- Analyze potential false positives (objects incorrectly identified) and false negatives (missed objects), define each clearly, and assess their likelihood and consequences in the context of urban and power infrastructure features.

- Check which city best matches airport placement, grid, highways, lights, and anomalies.

- Compare trade-offs of including vs. excluding objects or anomalies.

- Provide examples where relevant to clarify reasoning.

 

Steps:

1. Object Detection:

- Detect all major urban features, including:

- Buildings / downtown clusters

- Airports and airstrips

- Major highways and interchanges

- Bridges, rivers, and waterways

- Parks or open spaces

- Categorize each detected object: primary structure, transportation, natural feature, or other.

 

2. Anomaly Detection:

- Identify anomalies, such as unusual shapes, missing grid patterns, irregular lighting, or inconsistencies in infrastructure.

- Highlight anomalies with **red outlines**.

 

3. Image Simplification:

- Keep only essential shapes, layout, and primary objects.

- Remove minor details, textures, and noise.

 

4. Overlay Instructions:

- Overlay bright, high-contrast outlines on all detected objects.

- Ensure overlays are visually distinct and do not obscure the base image.

- Red outlines are reserved for anomalies.

 

5. Feature Description:

- Provide a concise list describing each outlined feature:

- Object category

- Approximate location

- Reason for detection

- Anomaly status if applicable

- Include discussion of potential false positives and false negatives with examples.

 

6. Summary and Reasoning:

- Summarize findings based on the most consistent reasoning across multiple paths.

- Note assumptions, uncertainties, likelihoods, and potential consequences.

- Provide a short conclusion highlighting the most reliable insights.

 

Constraints:

- Preserve original spatial layout and scale.

- Do not alter the base image except for overlays.

- Keep descriptions brief (1–2 sentences per object).

- Ensure overlays and highlights are easy to interpret.

 

AD #6: Optimized QA/QC Feature Mapping Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Less tokens than the prior version - 359 tokens/1933 characters in this version compared to 445 tokens/2389 characters in the prior version. It should be noted that one token roughly corresponds to about 4 characters of text (i.e., 100 tokens roughly equates to 75 words).

Task:

Analyze the uploaded image and produce a simplified, annotated version.

 

Instructions:

- Think carefully and reason step-by-step before producing your answer.

- Generate multiple reasoning paths, cross-check for consistency, and note assumptions or uncertainties.

- Analyze potential false positives (objects incorrectly identified) and false negatives (missed objects), define each clearly, and assess their likelihood and consequences.

- Compare trade-offs of including vs. excluding objects or anomalies.

- Provide examples where relevant.

- Summarize findings based on the most consistent reasoning.

 

Steps:

1. Object Detection:

- Detect all major objects and structural elements.

- Categorize each as: primary structure, transportation, natural feature, human, or other.

 

2. Anomaly Detection:

- Identify anomalies, unusual shapes, missing patterns, or irregular features.

- Highlight anomalies using red outlines.

 

3. Image Simplification:

- Keep only essential shapes, layout, and primary objects.

- Remove minor details, textures, or noise.

 

4. Overlay Application:

- Overlay bright outlines on all detected objects.

- Red outlines reserved for anomalies.

- Ensure overlays are clear and do not obscure the image.

 

5. Feature Description:

- Provide a concise list describing each outlined feature:

- Object category

- Approximate location

- Reason for detection

- Anomaly status if applicable

- Discuss potential false positives and false negatives with examples.

 

6. Summary and Reasoning:

- Summarize findings using the most consistent reasoning across paths.

- Note assumptions, uncertainties, likelihoods, and potential consequences.

- Provide a brief conclusion highlighting the most reliable insights.

 

Constraints:

- Preserve original spatial layout.

- Do not alter the base image except for overlays.

- Keep descriptions concise (1–2 sentences per object).

 

AD #7: Strategic Repetition Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Rearticulation via Reiteration using a Strategic Repetition mantra. As previously noted in AD #1, markdown in the form of **text** signifies bold; theoretically, the involved AI model should weight the associated instructions slightly higher.

Task:

Analyze the uploaded image and produce a simplified, annotated version.

 

Instructions:

- Think carefully and reason step-by-step before producing your answer. Repeat reasoning and cross-check at each stage.

- Generate multiple reasoning paths, cross-check for consistency, and note assumptions or uncertainties.

- Analyze potential false positives (objects incorrectly identified) and false negatives (missed objects), define each clearly, and assess likelihood and consequences. Reinforce this check after each detection step.

- Compare trade-offs of including vs. excluding objects or anomalies.

- Provide examples where relevant.

- Summarize findings based on the most consistent reasoning.

 

Steps:

1. Object Detection:

- Detect all major objects and structural elements.

- Categorize each as: primary structure, transportation, natural feature, human, or other.

- **Reinforce step-by-step reasoning here**: consider multiple detection paths and check for false positives/negatives.

 

2. Anomaly Detection:

- Identify anomalies, unusual shapes, missing patterns, or irregular features.

- Highlight anomalies using red outlines.

- **Double-check reasoning**: ensure anomalies are correctly identified and consistent with prior detection.

 

3. Image Simplification:

- Keep only essential shapes, layout, and primary objects.

- Remove minor details, textures, or noise.

- **Confirm simplification does not remove key objects or anomalies.**

 

4. Overlay Application:

- Overlay bright outlines on all detected objects.

- Red outlines reserved for anomalies.

- Ensure overlays are clear, visible, and do not obscure the image.

- **Cross-check overlays** against detected anomalies and objects; revise if necessary.

 

5. Feature Description:

- Provide a concise list describing each outlined feature:

- Object category

- Approximate location

- Reason for detection

- Anomaly status if applicable

- Discuss potential false positives and false negatives with examples.

- **Reinforce reasoning**: confirm descriptions are consistent with detected objects and anomalies.

 

6. Summary and Reasoning:

- Summarize findings using the most consistent reasoning across multiple paths.

- Note assumptions, uncertainties, likelihoods, and potential consequences.

- Provide a brief conclusion highlighting the most reliable insights.

- **Final cross-check**: ensure all reasoning, overlays, and description are consistent.

 

Constraints:

- Preserve original spatial layout.

- Do not alter the base image except for overlays.

- Keep descriptions concise (1–2 sentences per object).

- Ensure overlays and highlights are visually clear.

 

AD #8: Multi-Agent Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Utilizes various specialist vantage points in a multi-agent reasoning fashion.

Task:

Analyze the uploaded image and produce a simplified annotated version with object outlines, anomaly highlights, and structured reasoning.

 

Reasoning Requirements:

- Think carefully and reason step-by-step before producing the final answer.

- Generate multiple reasoning paths, cross-check them for consistency, and explicitly analyze potential false positives and false negatives.

- Define these clearly, provide examples, assess likelihood and consequences, compare trade-offs, note assumptions and uncertainties, and summarize conclusions based on the most consistent reasoning.

- Use hierarchical verification, triangulated verification, and a multi-agent reasoning structure as described below.

 

--------------------------------------------------

Agent Structure:

Agent 1 — Visual Detection Specialist

Detect all major objects and structural elements in the image.

Focus only on visual evidence: shapes, geometry, lighting patterns, spatial relationships.

 

Agent 2 — Infrastructure / Engineering Analyst

Classify detected objects from an infrastructure perspective.

Evaluate possible interpretations and generate multiple hypotheses.

 

Agent 3 — Dataset Verification Analyst

Cross-check detections against authoritative datasets when possible

(e.g., United States Geological Survey [USGS] or other geospatial sources).

 

Agent 4 — Anomaly and Risk Evaluator

Identify anomalies, unusual patterns, or inconsistencies.

Assess likelihood, consequences, and uncertainty.

Explicitly analyze false positives and false negatives.

 

Agent 5 — Final Synthesizer

Compare all reasoning paths and agent outputs.

Resolve inconsistencies and determine the most reliable interpretation.

 

--------------------------------------------------

Hierarchical Verification Process:

Level 1: Detect objects.

Level 2: Classify objects and generate hypotheses.

Level 3: Cross-check using external references and contextual reasoning.

Level 4: Evaluate anomalies and risk implications.

Level 5: Produce final synthesis and conclusions.

 

--------------------------------------------------

Image Processing Instructions:

1. Simplify the image while preserving essential shapes, layout, and primary objects.

2. Remove minor details, textures, and noise.

3. Overlay bright outlines on all detected objects.

4. Highlight anomalies using red outlines.

5. Ensure overlays do not obscure the base image.

 

--------------------------------------------------

Output Format:

1. Annotated image with overlays.

2. Object list including:

- object category

- approximate location

- reason for detection

- anomaly status

3. False positive / false negative analysis with examples.

4. Assumptions and uncertainties.

5. Final synthesis based on the most consistent reasoning.

 

AD #9: Multi-Pass Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Obligates the involved AI model to challenge its initial interpretation. By way of example, the various passes (via varied domains) facilitates discerning false positives while the alternative reasoning paths (via varied domains) facilitates discerning false negatives.

Task:

Analyze the uploaded image and generate a simplified annotated version highlighting detected objects and anomalies.

 

General Reasoning Instructions:

- Think carefully and reason step-by-step before producing the answer.

- Generate multiple independent reasoning paths, cross-check them for consistency, and select the most reliable interpretation.

- Explicitly analyze false positives and false negatives within the relevant engineering or infrastructure domain.

- Define each clearly, provide examples, assess likelihood and consequences, compare trade-offs, note assumptions and uncertainties, and summarize based on the most consistent reasoning.

 

--------------------------------------------------

PASS 1 — Visual Detection:

Detect all major objects and structural elements visible in the image.

Focus on:

- geometry

- edges

- shapes

- spatial relationships

- lighting and contrast

Produce a candidate object list.

 

--------------------------------------------------

PASS 2 — Independent Interpretation A:

Interpret the detected objects using infrastructure and engineering reasoning.

Generate hypotheses for each object’s possible function.

 

--------------------------------------------------

PASS 3 — Independent Interpretation B:

Re-evaluate the image independently from PASS 2.

Attempt alternative interpretations and challenge earlier assumptions.

Look specifically for:

- misidentified objects

- missing objects

- ambiguous structures.

 

--------------------------------------------------

PASS 4 — Dataset and Context Cross-Check:

Verify detections against trusted sources when possible such as:

- USGS infrastructure datasets

- geospatial context

- domain knowledge.

Flag inconsistencies or conflicts.

 

--------------------------------------------------

PASS 5 — Anomaly and Risk Analysis:

Identify unusual patterns or anomalies.

Explicitly evaluate:

False positives:

Objects detected but likely incorrect.

False negatives:

Objects that should exist but were not detected.

Assess:

- likelihood

- consequences

- operational risks

- uncertainty levels.

 

--------------------------------------------------

PASS 6 — Self-Consistency Comparison:

Compare reasoning outputs from PASS 2 and PASS 3.

Determine which interpretations are:

- consistent

- conflicting

- weakly supported.

Select the explanation supported by the strongest evidence.

 

--------------------------------------------------

Image Processing Instructions:

1. Create a simplified version of the image:

2. Preserve essential shapes and layout.

3. Remove minor details and textures.

4. Overlay bright outlines on detected objects.

5. Highlight anomalies using red outlines.

6. Keep the underlying image visible and unchanged.

 

--------------------------------------------------

Output Format:

1. Annotated simplified image with overlays.

2. Object detection table.

3. False positive analysis.

4. False negative analysis.

5. Assumptions and uncertainties.

6. Final synthesis based on the most consistent reasoning.

 

AD #10: Triangulated Evidence Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Utilizes evidence triangulation, which is comprised of visual evidence, contextual reasoning, and external verification.

Objective:

Analyze the provided image and generate a simplified annotated version highlighting essential structures, detected objects, and anomalies.

 

Core Reasoning Requirement:

- Think carefully and reason step-by-step before producing the answer.

- Generate multiple independent reasoning paths, compare them, and select the most consistent interpretation.

- Explicitly analyze false positives and false negatives within the relevant engineering or infrastructure domain.

- Define each clearly, provide examples, assess likelihood and consequences, compare trade-offs, note assumptions and uncertainties, and summarize conclusions based on the most consistent reasoning.

 

--------------------------------------------------

LAYER 1 — Visual Structure Extraction:

Identify primary visual features:

- edges

- shapes

- geometry

- spatial layout

- contrast patterns

- structural repetition

Produce an initial list of detected objects.

 

--------------------------------------------------

LAYER 2 — Multi-Agent Analysis:

Agent A — Visual Detection Specialist

Focus only on visual evidence. Identify candidate objects.

 

Agent B — Infrastructure / Engineering Analyst

Interpret objects using engineering knowledge.

 

Agent C — Context Analyst

Evaluate geographic and environmental context.

 

Agent D — Verification Analyst

Cross-check detections against trusted datasets or known patterns.

 

Agent E — Anomaly Detection Specialist

Identify unusual patterns and possible infrastructure anomalies.

 

--------------------------------------------------

LAYER 3 — Self-Consistency Reasoning:

Generate at least three independent reasoning paths:

- Reasoning Path A

- Reasoning Path B

- Reasoning Path C

 

Each path independently interprets the detected objects.

Compare the results and determine consensus findings.

 

--------------------------------------------------

LAYER 4 — Triangulated Evidence Scoring:

Evaluate each detected object using three evidence types:

- Visual evidence

- Contextual reasoning

- Dataset agreement

 

Assign confidence levels:

HIGH confidence = 3/3 evidence sources

MEDIUM confidence = 2/3 sources

LOW confidence = 1/3 source

 

--------------------------------------------------

LAYER 5 — Error Analysis:

Explicitly evaluate:

- False positives:

- Objects detected but likely incorrect.

- False negatives:

- Objects likely present but not detected.

 

For each case:

- explain reasoning

- provide examples

- estimate likelihood

- discuss consequences.

 

--------------------------------------------------

LAYER 6 — Image Simplification and Overlay:

Create a simplified version of the image:

1. Preserve essential layout and shapes.

2. Remove minor details and textures.

3. Outline all detected objects with bright overlays.

4. Highlight anomalies with red outlines.

5. Keep the base image visible and unchanged.

 

--------------------------------------------------

Final Output Structure:

1. Simplified annotated image with overlays

2. Detected object table:

- object type

- approximate location

- evidence sources

- confidence score

3. False positive analysis.

4. False negative analysis.

5. Assumptions and uncertainties.

6. Final synthesis based on the most consistent reasoning.

 

AD #11: Optimized Triangulated Evidence Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Less tokens than the prior version - 273 tokens/1425 characters in this version compared to 561 tokens/3280 characters in the prior version. It should be noted that one token roughly corresponds to about 4 characters of text (i.e., 100 tokens roughly equates to 75 words).

Task:

Analyze the image and produce a simplified annotated version with object outlines and anomaly highlights.

 

Reasoning:

Think carefully and reason step-by-step before answering.

Generate multiple reasoning paths and compare them before producing the final interpretation.

 

Process:

1. Detect major objects using visual cues:

shapes, edges, layout, spatial relationships.

2. Classify objects using infrastructure or contextual reasoning.

3. Generate at least two alternative interpretations and compare them.

4. Cross-check detections with contextual knowledge or known datasets.

5. Evaluate anomalies and explicitly analyze:

- false positives

- false negatives

Define each, give examples, and assess likelihood and consequences.

6. Assign confidence using triangulated evidence:

- Visual evidence

- Contextual reasoning

- Dataset agreement

 

Confidence levels:

High = 3 sources

Medium = 2 sources

Low = 1 source

 

Image Processing:

Create a simplified version of the image:

1. Preserve essential shapes and layout.

2. Remove minor details.

3. Outline detected objects with bright overlays.

4. Highlight anomalies in red.

5. Keep base image visible.

 

Output:

1. Annotated simplified image.

2. Object list with confidence levels.

3. False positive analysis.

4. False negative analysis.

5. Assumptions and uncertainties.

6. Final synthesis based on the most consistent reasoning.

 

AD #12: Uncertain Detections Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Obligates the involved AI model to identify assumptions and uncertain detections.

 

Task:

Analyze the image and generate a simplified annotated version with detected objects and anomaly highlights.

 

Reasoning:

Think carefully and reason step-by-step before answering.

Use multiple reasoning paths and compare them before producing the final interpretation.

 

--------------------------------------------------

Internal Verification Checklist:

Before producing the final answer, verify the following:

1. Object Detection

Did I detect all major shapes, structures, and layout elements?

2. Alternative Interpretations

Did I consider at least two possible interpretations for ambiguous objects

3. False Positive Check

Did I detect any objects that might not actually exist?

4. False Negative Check

Did I miss any objects that likely should be present?

5. Context Consistency

Are the detections consistent with the surrounding layout or environment?

6. Evidence Triangulation

Is each major conclusion supported by at least two of:

- visual evidence

- contextual reasoning

- known infrastructure patterns

7. Anomaly Review

- check for unusual patterns or inconsistencies?

8. Uncertainty Disclosure

- identify assumptions or uncertain detections?

Only produce the final answer after confirming all checklist items.

 

--------------------------------------------------

Image Processing:

Create a simplified image:

- preserve essential shapes and layout.

- remove minor details.

- outline detected objects with bright overlays.

- highlight anomalies in red.

- keep base image visible.

 

--------------------------------------------------

Output:

1. Annotated simplified image

2. Object detection list

3. False positive analysis

4. False negative analysis

5. Assumptions and uncertainties

6. Final synthesis based on the most consistent reasoning

 

AD #13: Optimized Uncertain Detections Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Less tokens than the prior version - 305 tokens/1732 characters in this version compared to 309 tokens/1771 characters in the prior version. It should be noted that one token roughly corresponds to about 4 characters of text (i.e., 100 tokens roughly equates to 75 words).

Task:

Analyze the image and produce a simplified annotated version with detected objects and anomaly highlights.

 

Reasoning:

Think carefully and reason step-by-step. Generate multiple independent interpretations and compare them before producing the final answer.

 

Pipeline:

1. Visual Detection

Identify major shapes, structures, and layout patterns.

2. Interpretation

Classify objects using engineering or contextual reasoning.

3. Alternative Reasoning

Generate at least two alternative interpretations and challenge earlier assumptions.

4. Evidence Triangulation

Evaluate each object using:

- visual evidence

- contextual reasoning

- dataset or pattern consistency.

 

Confidence:

High = 3 sources

Medium = 2

Low = 1

 

5. Error Analysis:

Explicitly evaluate:

- false positives

- false negatives

 

Define each, provide examples, and assess likelihood and consequences.

 

6. Anomaly Detection:

Identify unusual structures or patterns.

 

--------------------------------------------------

Internal Checklist:

Before answering verify:

- all major structures considered

- ambiguous objects reviewed

- false positives checked

- false negatives checked

- conclusions supported by evidence

- uncertainties noted.

 

--------------------------------------------------

Image output:

Create a simplified image:

- preserve essential layout and shapes.

- remove minor details.

- outline detected objects.

- highlight anomalies in red.

- keep base image visible.

 

--------------------------------------------------

Output:

1. Annotated simplified image.

2. Object list with confidence.

3. False positive analysis.

4. False negative analysis.

5. Assumptions and uncertainties.

6. Final synthesis.

 

AD #14: Visual-Attention Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Obligates the model to focus on each segment of the image, thereby minimizing the probability of missed objects while also increasing the probability of catching errors, via the checklist and self-consistency. In addition, processing each region sequentially reduces the likelihood of AI overload/hallucinations.

Task:

Analyze the uploaded image and produce a simplified annotated version with all detected objects and anomalies.

 

Reasoning:

Think carefully and reason step-by-step. Generate multiple independent reasoning paths and compare them before finalizing the answer.

 

--------------------------------------------------

Grid-based Visual Attention:

1. Divide the image into a logical grid of regions (e.g., 3x3 or 4x4).

2. Process each region sequentially:

- Detect shapes, objects, and patterns

- Classify detected objects

- Identify anomalies

3. For overlapping objects across regions, reconcile duplicates.

 

--------------------------------------------------

Evidence & Consistency:

- Triangulate evidence using:

- visual features in each region

- contextual reasoning

- cross-check with datasets (e.g., USGS, National Aeronautics and Space Administration [NASA], etc.)

- Evaluate confidence (High/Medium/Low) per object.

- Explicitly check for:

- false positives

- false negatives

- anomalous or unusual patterns

- Note uncertainties and assumptions.

 

--------------------------------------------------

Image Simplification & Overlay:

- Preserve essential layout and shapes

- Remove minor details

- Overlay detected objects with bright outlines

- Highlight anomalies in red

- Keep base image visible

 

--------------------------------------------------

Output:

1. Simplified annotated image with overlays

2. Object list with:

- category

- location

- confidence

- anomaly status

3. False positive analysis

4. False negative analysis

5. Assumptions and uncertainties

6. Final synthesis based on the most consistent reasoning

 

AD #15: Conjoined Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Conjoins varied Prompt Engineering (PE) techniques previously referenced.

Task:

Analyze the uploaded image and produce a simplified annotated version highlighting all detected objects, anomalies, and structural layouts.

 

Reasoning Instructions:

Think carefully and reason step-by-step. Generate multiple independent reasoning paths (self-consistency) and compare them before producing the final answer. Explicitly analyze false positives and false negatives, define each, provide examples, assess likelihood and consequences, note assumptions and uncertainties, and summarize conclusions based on the most consistent reasoning.

 

--------------------------------------------------

Grid-based Visual Attention:

Divide the image into a logical grid (e.g., 3x3 or 4x4 regions).

Sequentially process each region:

- Detect shapes, objects, and patterns

- Classify detected objects using engineering or contextual reasoning

- Identify anomalies

Reconcile overlapping or duplicate detections across regions.

 

--------------------------------------------------

Multi-Agent Analysis:

Agent A — Visual Detection Specialist: focuses on visual cues in each region.

 

Agent B — Infrastructure / Engineering Analyst: classifies and interprets.

 

Agent C — Context Analyst: evaluates environmental and spatial context.

 

Agent D — Dataset Verification Analyst: cross-checks with authoritative sources (e.g., USGS, NASA)

 

Agent E — Anomaly & Risk Evaluator: assesses unusual patterns, false positives/negatives, and uncertainties.

 

--------------------------------------------------

Triangulated Evidence & Confidence Scoring:

For each object, assign confidence based on:

- visual evidence

- contextual reasoning

- dataset or pattern agreement

 

Confidence Levels:

High = 3/3 sources, Medium = 2/3, Low = 1/3

 

--------------------------------------------------

Internal Verification Checklist:

Before producing the final answer, verify:

- all regions processed and major structures considered.

- ambiguous objects reviewed.

- false positives checked.

- false negatives checked.

- cross-region reconciliation completed.

- conclusions supported by triangulated evidence.

- anomalies highlighted.

- uncertainties and assumptions noted.

 

--------------------------------------------------

Image Simplification & Overlay:

- Preserve essential shapes and layout.

- Remove minor details.

- Overlay detected objects with bright outlines.

- Highlight anomalies in red.

- Keep base image visible.

 

--------------------------------------------------

Output Format:

1. Annotated simplified image with overlays.

2. Object detection table: category, location, confidence, anomaly status.

3. False positive analysis.

4. False negative analysis.

5. Assumptions and uncertainties.

6. Final synthesis based on the most consistent reasoning.

 

AD #16: Optimized Conjoined Prompt

(as referenced in Section IIIC. Experimentation)

Prospective Value-added Proposition: Less tokens than the prior version - 318 tokens/1783 characters in this version compared to 474 tokens/2778 characters in the prior version. It should be noted that one token roughly corresponds to about 4 characters of text (i.e., 100 tokens roughly equates to 75 words).

Task:

Analyze the image and produce a simplified annotated version with detected objects, anomalies, and layout.

 

Reasoning:

Think step-by-step. Generate multiple reasoning paths and select the most consistent interpretation. Analyze false positives and false negatives, note assumptions and uncertainties.

 

--------------------------------------------------

Grid-based Attention:

Divide image into regions (e.g., 3x3 or 4x4).

For each region:

- Detect shapes and objects

- Classify using engineering/context knowledge

- Identify anomalies

- Reconcile duplicates across regions.

 

--------------------------------------------------

Multi-Agent Reasoning:

Agent A: Visual Detection

Agent B: Engineering / Infrastructure

Agent C: Context Analysis

Agent D: Dataset Verification (e.g., USGS/NASA)

Agent E: Anomaly & Risk Evaluation

 

--------------------------------------------------

Evidence & Confidence:

Triangulate per object: visual, context, dataset.

Confidence: High=3, Medium=2, Low=1

 

--------------------------------------------------

Internal Checklist:

Before finalizing:

- All regions processed

- Ambiguous objects reviewed

- False positives/negatives checked

- Evidence supports conclusions

- Anomalies highlighted

- Uncertainties noted.

 

--------------------------------------------------

Image Simplification & Overlay:

- Preserve layout and shapes

- Remove minor details

- Outline objects with bright overlays

- Highlight anomalies in red

- Keep base image visible.

 

--------------------------------------------------

Output:

1. Annotated image

2. Object table: category, location, confidence, anomaly status

3. False positive analysis

4. False negative analysis

5. Assumptions and uncertainties

6. Final synthesis.