Robust Firearm Detection in Low-Light Surveillance Conditions Using YOLOv11 with Image Enhancement

Firearms, though essential for law enforcement and defense, pose serious threats when accessed by unauthorized individuals such as anti-social elements or mentally unstable persons. Several mass shooting incidents have occurred globally in recent years, particularly in public spaces such as schools, hospitals, and market areas [1, 2]. These events have intensified the need for advanced, real-time surveillance systems capable of identifying security threats, especially the presence of firearms.

Closed-Circuit Television (CCTV) systems have emerged as crucial components in public safety infrastructure, assisting in early threat detection and forensic investigation [3]. However, the efficiency of these systems remains limited due to the challenges posed by manual monitoring, especially under low-light or nighttime conditions. Surveillance effectiveness is significantly reduced in such environments, increasing the likelihood of missed detections [4].

Traditional firearm detection models, including those based on Faster R-CNN, SSD, and earlier YOLO variants, have demonstrated success in controlled lighting conditions [5-7]. However, their performance deteriorates under low illumination due to insufficient feature visibility, resulting in a high rate of false positives and missed detections [8-10]. Some recent studies have applied pre-processing techniques like brightness tuning and contrast normalization to mitigate these issues, but many of them compromise real-time performance or fail to generalize across varying lighting levels [11]. This research addresses a critical gap in the literature by focusing on firearm detection in low-light surveillance environments. It proposes an integrated solution combining a three-stage image enhancement pipeline adaptive gamma correction, Gaussian noise reduction, and min-max normalization with the state-of-the-art YOLOv11 detection model [12]. YOLOv11, presented at the YOLO Vision 2024 Conference (YV24), represents the latest evolution in the YOLO series and incorporates architectural innovations that significantly improve accuracy, detection speed, and computational efficiency [13, 14].

A recent advancement, such as the YOLOv7-DarkVision model [15], has demonstrated the merit of targeted pre-processing for object detection in dark conditions. Building upon this idea, the proposed system aims to improve real-time firearm detection accuracy by enhancing visual clarity before detection, rather than merely post-processing detection outputs.

The substantial advantages of using the proposed firearm detection system are:

• Enhanced Safety with Robust Surveillance: The system effectively detects firearms even in poorly lit environments, enabling timely alerts for law enforcement.
• Reduced Dependence on Manual Monitoring: The automated detection pipeline minimizes human fatigue and error.
• Improved Situational Awareness: Especially valuable in fenced zones or areas with limited lighting, it enhances visibility where artificial lighting is impractical.

In summary, this work contributes a computationally efficient, high-accuracy framework tailored for real-time firearm detection in dark environments. It fills a critical void in existing research and opens new avenues for the deployment of AI-enhanced surveillance in sensitive civilian and military applications.

Section II presents a comprehensive review of existing firearm detection methods and highlights gaps in performance under low-light conditions. Section III describes the dataset preparation process, including image sourcing, annotation, and the simulation of low-light environments through controlled darkening transformations. Section IV details the proposed methodology, including the selection and tuning of YOLOv11, the three-stage image enhancement pipeline, system architecture, and the end-to-end detection workflow. Section V provides a structured evaluation of the model, including quantitative results, enhancement effectiveness, visual performance, and comparative analysis with and without enhancement. Finally, Section VI concludes the study and outlines future research directions.

The dataset used in research work are legally and ethically collected from database platforms like Kaggle, Linksprite, VBS3, IMFDB, and UGR, and through independent internet searches. The details of image collection are provided in Table 2. A darkening transformation is used to create a low-light scenario on the collected images. To simulate a real-world scenario, various darkening factors have been used. The concept of darkening factors refers to the controlled reduction of pixel intensity values in an image to simulate darker environments or low lighting conditions. This is achieved by multiplying the pixel values of an image by a scalar factor, which reduces the brightness uniformly across all pixels. The darkening transformation is applied to the images after annotation and splitting. The Dataset is divided into training (70%), validation (15%), and testing (15%) sets. Figure 1 shows the sample images pairs from dataset with various darkening factor. In each image pair, the right side shows the darken image of the original image (left side). There are three darkening factor is considered for experimentation, to cover the complete dark environment scenario.

Table 2. Detail of collection of image dataset

1.jpg

Figure 1. Image samples pairs from dataset with various darkening factor

Having described the design and implementation of the proposed firearm detection system, the next section presents a comprehensive evaluation of its performance. The results are structured to demonstrate how the three-stage enhancement pipeline improves detection accuracy across varied lighting conditions. Both quantitative metrics and qualitative visualizations are used to validate the system's effectiveness in achieving robust, real-time firearm detection in low-light environments.

5.1 Evaluation metrics

To assess the performance of the proposed firearm detection model under varying lighting conditions, a standard set of classification and object detection metrics is used:

• True Positive (TP): The number of images correctly identified as containing firearms.
• False Positive (FP): The number of images incorrectly classified as containing firearms.
• True Negative (TN): The number of images correctly identified as not containing firearms.
• False Negative (FN): The number of images containing firearms but classified as not having them.

From these metrics proposed model Accuracy, Precision, Recall, F1-score, and mean Average Precision (mAP) are derived which are computed using Eqs. (6) to (9).

Precision (P): Indicates how many of the predicted firearm detections are actually correct.

$Precision (P R)=\frac{T P}{T P+F P}$                (6)

Recall (R): Measures how many actual firearm instances were correctly detected.

${Recall}(R C)=\frac{T P}{T P+F N}$              (7)

F1-Score: Harmonic mean of precision and recall, providing a balance between them.

$F1\ Score =2 \times \frac{P R \times R C}{P R+R C}$               (8)

Mean Average Precision (mAP@0.5): Represents the average precision computed at an IoU (Intersection over Union) threshold of 0.5, which evaluates the quality of bounding box predictions. Since this is a single-class (firearm) task, the mAP is directly computed over the firearm class only.

$m A P=\frac{1}{N} \sum_{i=1}^N A P_i$

$Since\ N=1(for\ Single\ class)$;

Hence,

$m A P=\sum_{i=1}^N A P_i$               (9)

Intersection over Union (IoU): Measures the overlap between the predicted bounding box and the ground truth box:

$I o U=\frac{ {Area}( {Intersection})}{ {Aera}({Union})}$          (10)

These metrics collectively provide a quantitative evaluation of how accurately and robustly the model detects firearms, especially in dark or low-visibility surveillance scenarios.

5.2 Experimental setup and dataset summary

To evaluate the proposed low-light firearm detection framework, a custom dataset consisting of 3,107 images was curated from publicly available sources including Kaggle, VBS3, IMFDB, UGR, and Linksprite. The dataset comprises images containing pistols, shotguns, and handguns, representing real-world surveillance scenarios. To simulate varying low-light environments, a darkening transformation was applied to original images by scaling down their pixel intensity values using three different darkening factors. This approach effectively mimics diverse night-time or poorly lit conditions typically encountered in outdoor and perimeter surveillance. The dataset was partitioned as follows:

• 70% for training,
• 15% for validation,
• 15% for testing.

Images were further categorized into three lighting condition groups based on Mean Pixel Intensity (MPI):

• S1 – Dark Conditions: MPI ≤ 20
• S2 – Moderate Darkness: 20 < MPI ≤ 55
• S3 – Normal Lighting: MPI > 55

For model performance comparison:

• A baseline YOLOv11n model (without enhancement) was trained and tested on all image categories.
• The proposed model integrated a three-stage enhancement pipeline (adaptive gamma correction, Gaussian filtering, and min-max normalization) prior to YOLOv11n detection.

Each experiment (S1, S2, S3) involved evaluating model accuracy, precision, recall, F1-score, and mAP@0.5 under these three lighting conditions. This stratified evaluation provides clear insight into how the proposed enhancements impact detection robustness in low-light surveillance.

5.3 Enhancement effectiveness analysis

The effectiveness of the proposed image enhancement pipeline was evaluated by comparing key image statistics before and after preprocessing across various lighting conditions. Table 4 summarizes representative results from the three experimental categories (S1: dark, S2: moderate-dark, S3: well-lit). For dark images (S1), gamma correction significantly improved brightness while preserving structural details.

For dark images (S1), gamma correction significantly improved brightness while preserving structural details. For example, images with a Mean Pixel Intensity (MPI) as low as 7.38 were brightened using a computed gamma value of 1.97, making firearm contours more visible to the detector. In these cases, standard deviation (representing contrast spread) was preserved, indicating that enhancement did not introduce over-smoothing.

Moderately dark images (S2) showed improved mean intensities with stable contrast metrics post-enhancement. In contrast, well-lit images (S3) bypassed gamma correction and normalization, undergoing only Gaussian filtering. This ensured natural lighting was preserved while still removing high-frequency noise.

Key Observations from Table 4:

• Gamma values adaptively decreased as MPI increased, ensuring controlled brightness enhancement.
• Mean pixel values remained stable after enhancement, validating the non-destructive nature of the pipeline.
• Standard deviation differences were minimal, confirming preservation of contrast details.
• I_min and I_max expansion in darker samples reflects improved dynamic range for detection.

These enhancements result in more consistent input quality, which directly supports improved model performance, particularly in low-light settings where traditional detection approaches tend to fail.

Table 4. Parameters observed before and after enhancement for selected images

5.4 Detection performance comparison

To assess the effectiveness of the proposed image enhancement pipeline, performance metrics were compared for the baseline YOLOv11n (without enhancement) and the enhanced YOLOv11n model across three lighting conditions S1 (dark), S2 (moderate-dark), and S3 (normal light). The key metrics considered are Accuracy, Precision, Recall, F1-Score, and mAP@0.5, as summarized in Table 5.

Key Observations from Table 5:

• In S1 (dark images), the proposed enhancement pipeline improved F1-score by 29.18%, recall by 28.05%, and mAP@0.5 by 35 points, demonstrating substantial gains in firearm detection accuracy under extremely low visibility.
• In S2 (moderate-dark conditions), improvements were also significant: a ~10% gain in F1-score and notable increases in recall and mAP.
• For S3 (well-lit conditions), only marginal improvement was observed. This validates the pipeline’s adaptive nature, where enhancement is selectively applied only to necessary cases, thus preserving original image quality.

Table 5. Comparative detection results for YOLOv11n with and without enhancement

5.5 Visual inspection of results

In addition to quantitative improvements, qualitative evaluation confirms the effectiveness of the proposed enhancement pipeline in enabling accurate firearm detection under challenging lighting conditions. Figures 5, 6, and 7 present representative samples from the three experimental categories:

Figure 5 showcases dark images (S1) undergoing full enhancement—gamma correction, Gaussian filtering, and normalization—followed by successful YOLOv11n detection. Step-by-step visualization of the proposed enhancement pipeline applied to dark surveillance images (S1). The sequence includes: original image, gamma-corrected image, Gaussian-filtered image, final normalized image, and YOLOv11n-based firearm detection. The enhancement stages significantly improve visibility and enable accurate localization of firearms in low-light conditions.

5.jpg

Figure 5. Detection results in dark images (S1)

Figure 6 includes moderately dark images (S2) where enhancement significantly improves visibility and localization, especially in cluttered or partially occluded scenes. Demonstration of enhancement and detection results for moderately dark images (S2). The proposed preprocessing pipeline improves feature visibility, enabling YOLOv11n to accurately detect firearms that are difficult to localize in the original frames. Red bounding boxes highlight correct detections with improved edge clarity.

6.jpg

Figure 6. Detection results in moderate-dark images (S2)

Figure 7 demonstrates well-lit images (S3) where gamma correction and normalization are skipped, and detection is performed directly after minimal filtering. Detection outputs for images captured under normal lighting conditions (S3). As per the adaptive pipeline, gamma correction and normalization are skipped, and only Gaussian filtering is applied. YOLOv11n performs accurately with minimal enhancement, confirming the efficiency of the adaptive strategy.

7.jpg

Figure 7. Detection results in well-lit images (S3)

Key Observations:

• In S1, firearm regions that were not visually distinguishable in the original images were successfully detected post-enhancement. This aligns with the improved F1-scores reported in Table 5 (from 36.5% to 65.68%).
• In S2, YOLOv11n consistently detected handguns across varying poses and backgrounds after enhancement, reflecting the precision and recall improvements.
• In S3, the model performed well with and without enhancement, validating the decision to bypass unnecessary processing in naturally lit scenes.

Figure 8 further supports the model’s robustness through a set of performance curves. The Precision-Confidence Curve shows a steady increase, reaching peak precision at a confidence threshold of 0.906. The Recall-Confidence Curve indicates high recall across most thresholds, with a gradual decline near 1.0. The Precision-Recall Curve confirms consistent performance with an mAP@0.5 of 0.895. The F1-Confidence Curve reveals the model's optimal detection threshold at 0.482, where the F1-score reaches 0.84. Together, these curves demonstrate reliable model behavior and help inform threshold selection for real-time deployment.

81.jpg

82.jpg

83.jpg

84.jpg

These visual inspections confirm that the proposed enhancement strategy directly contributes to addressing the primary research challenge: robust firearm detection under low-light surveillance conditions, where traditional detectors often fail.

5.6 Summary of findings and future directions

The experimental results, supported by both quantitative metrics and visual inspection, demonstrate that the proposed enhancement-integrated YOLOv11n framework significantly improves firearm detection performance in low-light surveillance conditions. This addresses a critical challenge faced by traditional real-time detectors, which tend to suffer from degraded accuracy in visually constrained environments. Compared to the baseline YOLOv11n, the enhanced model achieved:

• A 29.18% absolute increase in F1-score and 35-point mAP@0.5 boost in the darkest test cases (S1),
• A ~10% gain in recall and precision in moderately dark conditions (S2),
• Near-equivalent performance in well-lit settings (S3), validating the adaptive enhancement logic.

These findings are consistent with and build upon recent literature such as Yadav et al. [38], where YOLOv7-DarkVision was used to improve detection in dark environments. However, the current approach takes a step further by:

• Introducing an adaptive, conditional enhancement pipeline,
• Leveraging YOLOv11’s architectural efficiencies, and
• Demonstrating robustness across multiple lighting ranges with practical computational cost.

Future Directions: To further enhance the applicability and scalability of this work, the following research avenues are suggested:

• Model Generalization to Other Threat Classes: Extend the system to detect other concealed or handheld threats (e.g., knives, explosives) in dark environments using multi-class training.
• Hardware Optimization and Edge Deployment: Port the model to edge computing devices (e.g., Jetson Nano, Coral TPU) to enable deployment in bandwidth- or power-constrained environments.
• Integration with Tracking and Alert Systems: Combine detection with temporal tracking and automated alert mechanisms to create a complete real-time threat monitoring solution.
• Federated Learning for Privacy-Sensitive Environments: Implement decentralized training strategies for sensitive locations (schools, defense zones) to enable secure, privacy-aware model updates.

By aligning detection robustness with real-time performance, this work contributes toward the development of intelligent, scalable firearm detection systems for next-generation surveillance infrastructures.