HOTSPOT-YOLO for Thermal Anomaly Detection

• HOTSPOT-YOLO is a specialized YOLO-derived framework that integrates attention and multi-scale aggregation to detect small, low-contrast thermal anomalies.
• The model employs an efficient backbone with depthwise separable convolutions and Squeeze-and-Excitation blocks, achieving higher mAP with reduced computational cost.
• HOTSPOT-YOLO has spurred broader design extensions, including hypergraph-based contextual modeling and applications in IC layout hotspot detection.

Searching arXiv for HOTSPOT-YOLO and related hotspot-detection YOLO papers to ground the article. HOTSPOT‑YOLO is a YOLO‑derived object detection framework associated with at least two distinct research usages on a domain‑specialized detector for thermal anomaly detection in drone‑based solar photovoltaic inspections, and a broader conceptual extension in which hypergraph‑based contextual modeling is proposed as a possible blueprint for hotspot‑oriented detection systems (Dhimish, 26 Aug 2025, Abid et al., 3 Jun 2026). In the published arXiv literature, the name is used most concretely for a lightweight, attention‑driven YOLOv11‑based model for detecting thermal hotspots and defective modules in infrared imagery of PV fields, where the central challenge is reliable localization of small, low‑contrast anomalies under real‑time deployment constraints (Dhimish, 26 Aug 2025). Separately, the term “hotspot” is also used in IC layout verification, where YOLO‑based systems detect lithographic layout hotspots through object detection on rasterized layout imagery, although that work does not itself name the system HOTSPOT‑YOLO (Wu et al., 2024). Taken together, these strands position HOTSPOT‑YOLO as a family resemblance rather than a single canonical architecture: a hotspot‑centric adaptation of YOLO in which backbone design, feature aggregation, and contextual augmentation are specialized to the structure of hotspot phenomena.

1. Terminological scope and research context

The clearest arXiv usage of the name appears in “HOTSPOT‑YOLO: A Lightweight Deep Learning Attention‑Driven Model for Detecting Thermal Anomalies in Drone‑Based Solar Photovoltaic Inspections” (Dhimish, 26 Aug 2025). In that work, HOTSPOT‑YOLO denotes a detector specifically designed for drone‑captured infrared images of solar photovoltaic plants, with the primary class labeled “anomaly” and encompassing thermal hotspots and related defects. The problem setting is operational rather than generic: large PV fields contain thousands of modules, faults manifest as small temperature differences, manual inspection is slow and error‑prone, and standard detectors are not tailored to thermal imagery with low contrast, noise, and environmental variability (Dhimish, 26 Aug 2025).

A second, more speculative line arises from “HYolo: An Intelligent IoT‑Based Object Detection System Using Hypergraph Learning” (Abid et al., 3 Jun 2026). That paper does not implement a model named HOTSPOT‑YOLO, but it explicitly discusses how its hypergraph‑enhanced YOLOv8‑N architecture could inspire a future “HOTSPOT‑YOLO” in which hotspot regions, sensor regions, or correlated high‑activity areas are treated as hypergraph vertices and linked by hyperedges. This suggests a broader interpretation of HOTSPOT‑YOLO as a hotspot‑prioritizing YOLO framework with explicit high‑order contextual reasoning, although this remains a proposed extension rather than a reported implementation (Abid et al., 3 Jun 2026).

A third relevant usage of “hotspot” appears in electronic design automation. “Enhancing Layout Hotspot Detection Efficiency with YOLOv8 and PCA‑Guided Augmentation” applies YOLOv8n to detect multiple lithographic hotspots in IC layout images, augmented by a PCA‑derived channel encoding local layout complexity (Wu et al., 2024). This is not HOTSPOT‑YOLO by name, but it establishes that the “hotspot + YOLO” formulation extends beyond thermal imaging into manufacturing verification.

2. HOTSPOT‑YOLO for photovoltaic thermal anomaly detection

In the solar PV setting, HOTSPOT‑YOLO is an enhanced version of Ultralytics YOLO11 adapted for thermal imagery (Dhimish, 26 Aug 2025). The detector retains the YOLOv11 detection head and single‑stage formulation but changes the feature extractor and introduces attention mechanisms intended to improve sensitivity to small and subtle thermal anomalies while maintaining low computational cost.

The application context is tightly defined. Drone thermal images are resized and normalized, and the target anomalies include hotspots, defective modules, and mismatched or degraded cells as they appear in infrared inspections. The framework is designed to address four specific constraints reported in the paper: small and subtle targets, low contrast and high noise, varying environmental conditions, and real‑time inference on GPU or edge devices (Dhimish, 26 Aug 2025). The main anomaly class is represented as a single “anomaly” label, although the paper states that the framework is suitable for other PV thermal faults as well.

The high‑level data flow begins with a drone thermal image of size $640 \times 640 \times 3$, followed by resizing and normalization, backbone feature extraction, feature aggregation, and YOLOv11 multi‑scale detection, with post‑processing by Non‑Maximum Suppression (Dhimish, 26 Aug 2025). For each grid cell, the model predicts

$y = \{b, c, p\},$

where $b=(x,y,w,h)$ is the bounding box, $c$ is the class label, and $p$ is the confidence. Final detections are filtered through

$$B_{\text{final}} = \text{NMS}(B, \text{IoU threshold}).$$

This architecture is described as domain‑specialized rather than wholly novel: it is still a YOLO detector, but one reconfigured for PV thermography through a lightweight EfficientNet‑style backbone, Squeeze‑and‑Excitation channel attention, and a dedicated multi‑scale feature aggregation block (Dhimish, 26 Aug 2025).

3. Backbone design, attention, and multi‑scale aggregation

The backbone in HOTSPOT‑YOLO is an EfficientNet‑like lightweight CNN using depthwise separable convolutions (Dhimish, 26 Aug 2025). The stated rationale is the accuracy‑parameter tradeoff, suitability to noisy solar panel data, and low FLOPs for real‑time inference. The thermal image is resized from $640 \times 640$ to a backbone input of $224 \times 224$. The initial convolution maps $224 \times 224 \times 3$ to $112 \times 112 \times 32$, after which depthwise and pointwise convolution blocks progressively produce shallow, intermediate, and deep features of sizes $y = \{b, c, p\},$0, $y = \{b, c, p\},$1, and $y = \{b, c, p\},$2 (Dhimish, 26 Aug 2025).

The paper writes the depthwise separable convolutional update as

$y = \{b, c, p\},$3

with $y = \{b, c, p\},$4 as ReLU. It also reports the standard FLOPs comparison between ordinary convolution and depthwise separable convolution, emphasizing that the latter substantially reduces cost when $y = \{b, c, p\},$5 (Dhimish, 26 Aug 2025). This reduction is central to the model’s lightweight characterization.

Attention is supplied by Squeeze‑and‑Excitation blocks integrated into each backbone stage (Dhimish, 26 Aug 2025). Given $y = \{b, c, p\},$6, global average pooling computes

$y = \{b, c, p\},$7

and the channel weights are produced by a two‑layer MLP with ReLU and sigmoid. In the paper’s notation, the recalibrated feature map is

$y = \{b, c, p\},$8

The stated role of SE is to amplify channels corresponding to thermally significant patterns and suppress less informative responses, especially when hotspots are weak signals against noisy thermal backgrounds (Dhimish, 26 Aug 2025).

Feature aggregation fuses low‑, mid‑, and high‑level features via element‑wise addition and a $y = \{b, c, p\},$9 convolution to produce a unified multi‑scale feature map $b=(x,y,w,h)$0 (Dhimish, 26 Aug 2025). The detection head then operates at three scales for small, medium, and large objects. This combination of lightweight convolution, channel attention, and multi‑scale fusion is the paper’s main architectural basis for improved hotspot sensitivity.

4. Detection head, loss formulation, and training protocol

HOTSPOT‑YOLO uses the YOLOv11 detection head with multi‑scale detection at three feature map sizes and anchor‑based prediction with auto‑anchor learning (Dhimish, 26 Aug 2025). For each cell and scale, the output again takes the form

$b=(x,y,w,h)$1

where $b=(x,y,w,h)$2 is defined as objectness multiplied by class probability.

The total loss is reported as the standard YOLOv11 decomposition

$b=(x,y,w,h)$3

where $b=(x,y,w,h)$4 is a bounding box regression loss based on IoU between predicted and ground‑truth boxes, $b=(x,y,w,h)$5 is a classification loss, and $b=(x,y,w,h)$6 is the objectness or confidence loss (Dhimish, 26 Aug 2025). IoU is defined as

$b=(x,y,w,h)$7

The paper notes that the exact regression variant is not explicitly detailed; it states that standard YOLOv11 typically uses CIoU or SIoU, so any stronger claim would be inferential rather than directly reported (Dhimish, 26 Aug 2025).

Training data come from Roboflow thermal PV anomaly datasets (Dhimish, 26 Aug 2025). The details reported include one primary anomaly class, training sets described as 3600–6000 images, validation sets of 470–720 images, and a test set of 450 images, all normalized and resized to $b=(x,y,w,h)$8 pixels. Augmentation includes horizontal flip and random crop, with robustness experiments additionally considering brightness/contrast variation and grayscale conversion. Optimization uses Adam with initial learning rate $b=(x,y,w,h)$9, momentum $c$0, weight decay $c$1, cosine decay, batch size 16, and up to 200 epochs, with convergence around 120–125 epochs on an NVIDIA A100 GPU in Google Colab (Dhimish, 26 Aug 2025).

Mean Average Precision is defined as

$c$2

with $c$3 for the anomaly class, and the paper reports steady mAP growth with saturation near 125 epochs (Dhimish, 26 Aug 2025).

5. Empirical performance and robustness characteristics

The central reported performance figures for HOTSPOT‑YOLO are an [email protected] of 90.8%, 36.10M parameters, 25.53 GFLOPs, and 25.22 ms inference time per $c$4 image (Dhimish, 26 Aug 2025). Relative to the YOLOv11m baseline, the paper reports a +5.9% mAP improvement, a reduction of 2.12M parameters, lower FLOPs, and faster inference (Dhimish, 26 Aug 2025).

The benchmark comparison reported in the paper is as follows:

Model	mAP (%)	Params/FLOPs/Time
Cascade RCNN	74.2	68.93M / 80.15G / 25.57 ms
CenterNet	68.5	28.89M / 142.13G / 27.62 ms
Faster RCNN	75.8	41.12M / 41.12G / 27.43 ms
YOLOv5m	81.4	18.12M / 15.83G / 22.35 ms
YOLOv9m	80.3	25.30M / 105.2G / 23.76 ms
YOLOv11m	84.9	38.22M / 29.61G / 29.61 ms
HOTSPOT‑YOLO	90.8	36.10M / 25.53G / 25.22 ms

The paper interprets these results as showing the highest mAP among the tested detectors, while remaining lighter than YOLOv11m and substantially lighter than the RCNN baselines in both parameters and FLOPs (Dhimish, 26 Aug 2025). Training curves are reported to show smoother and faster convergence than YOLOv5m, YOLOv9m, and YOLOv11m.

Robustness experiments examine several non‑ideal imaging conditions (Dhimish, 26 Aug 2025). Under high irradiance and uniform heating, HOTSPOT‑YOLO reportedly isolates true anomalous modules with 91–94% confidence. In images containing irrelevant hot surfaces and temperature scale bars, the model focuses on module hotspots with 78–79% confidence. Under brightness and contrast reductions of 40%, detections remain near 73% confidence, while grayscale conversion produces a modest drop to 68% and 67%. Strong blur yields 53–71% confidence, and moderate blur 79–88%. On ground‑based thermal images, despite being trained only on drone data, the model achieves 63–82% confidence in wide‑angle views and up to 85% in close‑ups, with some distant or edge anomalies dropping to 28% due to domain shift (Dhimish, 26 Aug 2025).

These findings are framed in the paper as evidence of robustness to irradiance changes, contrast variation, blur, noise, and moderate drone‑to‑ground domain shift (Dhimish, 26 Aug 2025). A plausible implication is that the combined effect of SE attention and multi‑scale aggregation improves anomaly salience even when raw thermal contrast is degraded.

6. HOTSPOT‑YOLO as a broader design pattern

Although the PV detector is the only arXiv paper in the provided corpus to use the name HOTSPOT‑YOLO explicitly, related work reveals a broader design space. In HYolo, hotspot‑focused detection is discussed as a potential future extension of hypergraph‑enhanced YOLO (Abid et al., 3 Jun 2026). That model is based on YOLOv8‑N and embeds hypergraph learning into the neck through HyperC2Net and HyperConv, with the goal of modeling high‑order feature relationships that ordinary YOLO necks do not capture explicitly.

The hypergraph is defined as

$c$5

with incidence matrix

$c$6

and distance‑based hyperedge construction

$c$7

using an empirically chosen threshold $c$8 (Abid et al., 3 Jun 2026). Hypergraph propagation is performed by

$c$9

The paper proposes that a future “HOTSPOT‑YOLO” could define hotspot proposals, sensor regions, or activity clusters as vertices and use hyperedges to connect spatially, semantically, or temporally correlated hotspots (Abid et al., 3 Jun 2026).

This usage does not describe a deployed system, but it is significant because it expands “hotspot” from a local thermal defect to a more general object‑context unit. In that sense, HOTSPOT‑YOLO can also denote a hotspot‑centric neck design in which contextual structure, not merely per‑region appearance, drives prediction. The HYolo results on COCO report that Hyper‑YOLO‑N improves over baseline YOLOv8‑N from 0.155 to 0.175 mAP@50 and from 0.085 to 0.105 [email protected]:0.95, while reducing box loss from 1.60 to 1.45 and increasing peak F1‑score from 0.62 to 0.72 (Abid et al., 3 Jun 2026). This suggests, though does not prove for PV inspection specifically, that explicit higher‑order context modeling may be a productive extension path for future hotspot detectors.

7. Related hotspot detection paradigms and limitations

The most important adjacent hotspot‑detection literature in the provided material is the IC layout hotspot work using YOLOv8n with PCA‑guided augmentation (Wu et al., 2024). There, hotspots are lithographic failure patterns in rasterized layout images, and the key innovation is an extra image channel encoding local layout complexity derived from PCA. The method achieves approximately 83% precision and 86% recall while keeping the false alarm rate below 7.4%, and it improves never‑seen‑before hotspot detection by about 10% (Wu et al., 2024). The relevance to HOTSPOT‑YOLO is conceptual: hotspot detection often benefits from specialized auxiliary representations rather than naïve reuse of a generic YOLO backbone.

In the IC layout study, the detector is YOLOv8n with about 3.2M parameters, trained on $p$0 layout mosaics derived from ICCAD‑2019. Reported inference time is approximately 5 ms per image on an RTX 2080, excluding PCA feature extraction, while CPU PCA extraction costs about 10 ms per clip (Wu et al., 2024). The work shows that hotspot detection frequently involves small, visually similar patterns and benefits from explicitly engineered context or complexity channels.

Across these HOTSPOT‑YOLO‑related variants, several limitations recur. In the PV formulation, limitations include domain shift to ground‑based images, sensitivity to thermal camera calibration and environmental conditions, reliance on a single thermal modality, limited class granularity, and insufficient quantification under extreme weather conditions (Dhimish, 26 Aug 2025). In the hypergraph formulation, computational overhead and sensitivity to hypergraph construction parameters, especially $p$1, are identified as constraints (Abid et al., 3 Jun 2026). In the IC layout formulation, precision degrades under scale mismatch and PCA feature extraction can become a bottleneck (Wu et al., 2024).

Future work proposed across the papers converges on several themes: edge deployment optimization through compression or quantization, broader and more diverse training data, adaptive or learned contextual structures, and richer multimodal inputs such as RGB‑thermal fusion or multi‑sensor integration (Dhimish, 26 Aug 2025, Abid et al., 3 Jun 2026). This suggests that HOTSPOT‑YOLO is likely to evolve not as a single frozen model name, but as a recurrent pattern in which YOLO is specialized for hotspot phenomena through lightweight backbones, attention or context modules, and domain‑specific representations.