Solar Panel Research-Object detection Training and Classification

🤝 Community Help
1

Research Goal

I am an undergraduate researcher working with false-color thermal images of solar panels. My three defect classes are:

  • Bypass_Diode
  • Cell_Fault
  • Thermal_Hotspot

My current Roboflow dataset is an Object Detection project with bounding-box annotations. Some images contain one defect, some contain multiple instances of the same defect, and some images contain multiple different defect classes.

My goal is to use the cleaned dataset to compare multi-label image classification and YOLO object detection under similar experimental conditions. For classification I am considering lightweight models such as MobileNetV3-Small, EfficientNet-Lite0, and ShuffleNetV2. For object detection I plan to compare lightweight YOLO models. Eventually I am interested in edge/drone deployment.

Question 1 — What am I seeing in Roboflow?

When I open an image in my Object Detection project, I see colored bounding boxes around defects with class names such as Cell_Fault or Thermal_Hotspot.

Are these boxes the ground-truth annotations that are used to teach the YOLO model, or are they output/predictions generated by an AI model?

During YOLO training, what exactly does the model receive from Roboflow? Does it receive the original image pixels plus the class IDs and numerical bounding-box coordinates, while the colored boxes I see in the Roboflow interface are only a visualization of those annotations?

Also, after training, where in Roboflow can I view the model’s predicted bounding boxes so I can clearly compare ground truth vs model output?

Question 2 — Using the same Object Detection dataset for Multi-Label Classification

I want to compare YOLO object detection against full-image multi-label classification using essentially the same source images.

For example, if one image has:

Cell_Fault × 3 bounding boxes
Thermal_Hotspot × 1 bounding box

I want YOLO to retain all four boxes, but I want the classification version of the same full image to become:

Bypass_Diode = 0
Cell_Fault = 1
Thermal_Hotspot = 1

I do not want to crop each bounding box into a separate classification image.

Is exporting my cleaned Object Detection dataset as Multi-Label Classification CSV the correct Roboflow workflow for accomplishing this?

Question 3 — How should I set up a fair comparison?

I want to create one cleaned master Object Detection dataset and then branch it into:

Clean Object Detection Master → YOLO dataset

and

Clean Object Detection Master → Full-image Multi-Label Classification dataset

What is the recommended way in Roboflow to make sure the same source images and same Train/Validation/Test assignments are used in both experiments?

Should I clean and correct the Object Detection master first, generate the final Train/Validation/Test split, and then export that same dataset version separately for YOLO and Multi-Label Classification?

Question 4 — Comparing the results

Because classification and object detection normally use different metrics, what would be a reasonable way to compare them fairly?

My idea is to evaluate the multi-label classifier using image-level class presence, and also convert YOLO predictions into image-level presence (for example, if YOLO detects at least one Cell_Fault, count Cell_Fault as present). Then I could compare both approaches using the same image-level precision, recall, and F1 metrics.

YOLO would additionally be evaluated with detection metrics such as mAP and localization performance.

Does this experimental setup make sense, or is there a better way you would recommend structuring the comparison?

Question 5 — Composite thermal images

For images containing two or three different defect classes, is there anything special I should do in Roboflow before converting the Object Detection dataset to Multi-Label Classification?

I want to make sure composite images retain all of their image-level class labels during conversion. Some of my thermal solar-panel images contain multiple defect types in the same full image, such as Cell_Fault + Thermal_Hotspot, or possibly all three classes (Bypass_Diode + Cell_Fault + Thermal_Hotspot).

I am currently considering full-image multi-label classification, where one image can receive multiple labels, while keeping the original bounding-box annotations for the YOLO object-detection experiment.

Besides full-image multi-label classification, what other approaches would you recommend considering for composite images?

For example, would any of these approaches be appropriate?

  • Crop the annotated bounding boxes and perform classification on individual defect regions.
  • Use full-image multi-label classification.
  • Keep composite images primarily for object detection.
  • Use a two-stage system where YOLO first detects/crops defect regions and a classifier then classifies the detected regions.
  • Use another approach available in Roboflow that I may not be considering.

For a research comparison between classification and YOLO object detection, which approaches are technically reasonable to investigate, and what are the advantages or limitations of each?

I also want to eventually deploy the system on an edge device or drone, so computational cost, inference speed, model size, and ability to handle multiple defects are important considerations.

2

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.