A Machine Learning-Based Tool for Indoor Lighting Compliance and Energy Optimization

The system developed in this study represent a real-time indoor illuminance estimation tool that utilizes a smartphone's onboard camera with a trained machine learning model. Illuminance, or the total luminous flux per unit area falling on a surface, is quantified in lux (lx). The mathematical representation of illuminance is given ass:

$E=\frac{\Phi}{A}$                                  (1)

where, E is illuminance in lux, $\Phi$ is luminous flux in lumens, and A is area in square meters [12]. Illuminance serves as a quantitative metric for assessing the lighting adequacy of a surface, which is crucial for evaluating visual comfort and lighting quality in indoor environments.

The proposed system consists of four components: (i) a user-friendly mobile application interface, (ii) a machine learning illuminance prediction model, (iii) a heat map visualization module to represent spatial light distribution, and (iv) a recommendation algorithm for assessing compliance with recognized lighting standards. The complete workflow of the platform is illustrated in Figure 1.

501.png

Figure 1. Platform architecture and data flow

At the front end, the application enables the user to select a specific indoor environment from a predefined list of use cases (e.g., residential, classroom, retail). Once the analysis is initiated, the smartphone's rear camera activates and captures a frame of the current scene. Images were captured using a Xiaomi Redmi Note 8 smartphone, equipped with a 48MP camera (f/1.8, 1/2.0", 0.8µm) at a 524×324 pixels resolution prior to processing and model training. This resolution was empirically found to offer a balance between spatial detail and computational efficiency and visual aesthetic.

The captured image is subsequently divided into small 2×2 pixel patches, providing a fine-grained assessment of lighting distribution while maintaining low computational overhead. Each patch is analyzed based on the mean intensity values of its red, green, and blue (RGB) channels, which serve as the input features for the illuminance prediction model.

The model selected in our study was implemented using TensorFlow.js that allows machine learning models to make inference locally on the device without the need for external servers which enhances both security and offline accessibility. For each patch, the model estimates the corresponding lux value. These values are then stored into a two-dimensional illuminance matrix which represent the predicted light intensity across the captured scene.

Viridis color map is then used to create a color-coded heat map overlay to depict the spatial distribution of illuminance levels throughout the scene. The overlay is rendered on top of the original image via the use of canvas blending techniques, allowing dynamic, intuitive, and real-time visualization.

Simultaneously, the platform also computes the general average illuminance value of the whole frame and compares it against the recommended lux levels defined for the selected indoor setting. Based on this comparison, the application app categorizes regions as underlit, adequately lit, or overlit, and delivers contextual lighting recommendations to the user.

This methodology enables intuitive interpretation of lighting adequacy by non-expert users, removing the need for specialized instrumentation or technical knowledge. The platform's emphasis on accessibility and real-time responsiveness supports its applicability in diverse settings, including education, healthcare, office, and residential environments.

2.1 Data collection and labelling

To develop a robust and accurate predictive model for real-time indoor illuminance estimation using smartphone imagery, a dedicated dataset was constructed under controlled experimental conditions. The objective was to establish a quantitative relationship between visual features extracted from image patches and their corresponding ground truth illuminance values (in lux), as measured by a calibrated physical sensor.

The data acquisition process was carried out in a scaled indoor mock-up environment designed to replicate typical residential and office lighting conditions, as shown in Figure 2. The lighting environment consisted of both natural daylight whose intensity varied over time and artificial lighting provided by a dimmable 5730 white LED strip.

502.png

Figure 2. Scaled mock-up for data collection process

The artificial lighting intensity modulated using a potentiometer interfaced with an Arduino microcontroller via Pulse Width Modulation (PWM) signals, enabling precise and continuous control of the lighting output. This setup facilitated the simulation of diverse real-world lighting scenarios across different times of day, enhancing the robustness of the dataset for training and evaluation purposes.

To collect paired data points, a smartphone was mounted in a fixed and stable position to periodically capture images of a predefined target area within the indoor mock-up environment. The image resolution was fixed, and camera parameters including white balance, ISO, exposure, and focus were manually locked to reduce the influence of automatic software enhancements and ensure consistency and reproducibility throughout the dataset.

Images were captured at a resolution of 524×324 pixels before patch extraction. Simultaneously, a BH1750 ambient light sensor was placed at the center of the imaged scene and connected to the Arduino platform, the BH1750 is a digital light sensor capable of measuring illuminance from 1 to 65535 lux with an accuracy of ±0.5%, and was chosen for its reliability and ease of integration via the I2C protocol. The complete hardware configuration is illustrated in Figure 3.

503.png

Figure 3. Scaled room interior data collection process (a), custom-built illuminance sensor circuit (b)

For each captured image, a synchronized illuminance reading was recorded using timestamp alignment, enabling a precise one-to-one mapping between each image and its respective lux measurement.

To eliminate the effect of surface reflectance and ensure that the lux measurements reflected true light intensity, the scene’s surface was covered with a uniform layer of neutral gray matte material (18% reflectance). This is a widely adopted reference surface in professional photographic and lighting calibration, and it helped minimize bias in RGB values caused by glossy, reflective, or dark absorptive textures, while also reducing specular noise in the captured images.

For each collected image a 2×2 pixel patch was segmented and extracted, a resolution selected to provide a trade-off between spatial granularity and computational efficiency. For each patch, four features extracted and stored for training:

·Mean intensity of the red channel (R_avg).

·Mean intensity of the green channel (G_avg).

·Mean intensity of the blue channel (B_avg).

·The average grayscale intensity (GS_avg).

The grayscale intensity was calculated using the standard luminance transformation, defined as follows:

$G S=0.2989 * R+0.5870 * G+0.1140 * B$                              (2)

where, R, G, and B are red, green, and blue channel intensities, respectively.

These four normalized values were normalized to a [0–1] range and used as input to the regression models, while the corresponding illuminance (ILS), measured in lux, by the BH1750 sensor, served as the target output label. Although the lux sensor captures only a single-point reading, the captured scene was carefully composed and uniformly illuminated to ensure that the recorded value accurately represented the lighting condition of the imaged region.

Data acquisition was performed at one-hour intervals throughout an entire calendar year, capturing a wide range of lighting scenarios from low-light conditions in the early morning and high-intensity illumination at midday, to artificially lit environments during the evening.

The final curated dataset comprises 2,650 entries, each representing a distinct 2×2 patch under specific lighting conditions, paired with a corresponding ground-truth illuminance value (in lux), measured using a BH1750 digital light sensor. Each data entry includes R_avg, G_avg, B_avg, GS_avg as features and ILS as target output.

Illuminance values in the dataset span a wide range, from 0 lux to over 1,200 lux, successfully capturing a representative distribution of low, moderate, and high lighting conditions typically encountered in indoor spaces.

To further analyze the internal coherence of the dataset, a 3D scatter plot (Figure 4) was generated to visualize the relationship between G_avg, GS_avg, and ILS. The resulting plot reveals a smooth, monotonically increasing surface, with no evident anomalies, thereby supporting the internal consistency and reliability of the collected dataset.

504.png

Figure 4. 3D scatter plot_ G_avg vs GS_avg vs ILS

2.2 Model design and training

The indoor illuminance estimation from image-derived features constitutes a regression task characterized by a continuous output space. The principal objective of this step is to construct predictive models capable of accurately estimating illuminance from low-resolution 2×2 color pixel patches. To achieve this, three machine learning models were selected, trained, and evaluated:

·Multi-Layer Perceptron (MLP) Regressor.

·Random Forest Regressor (RFR).

·Extreme Gradient Boosting (XGBoost) Regressor.

These models were chosen based on their established success in image-based regression tasks, and their ability to model complex nonlinear relationships between input features and target outputs.

2.2.1 Multi-Layer Perceptron regressor

The Multi-Layer Perceptron (MLP) is a fully connected feedforward neural network widely used for regression and classification due to its ability to model nonlinear input-output relationships. It consists of an input layer, one or more hidden layers, and an output layer, where each neuron computes a weighted sum of inputs, adds a bias, and applies an activation function [13]. The operation of a single neuron can be expressed as:

$a^{(l)}=\Phi\left(\sum \omega_{i j}^{(l)} a_j^{(l-1)} b_i^{(l)}\right)$                             (3)

where, $a^{(l)}$ is the activation function of the neuron in layer $l$, $\omega_{i j}^{(l)}$ is the weight connecting neuron $j$ in layer $l$-$1$ to neuron $i$ in layer $l$, $b_i^{(l)}$ is the bias term, and $\Phi$ is the activation function, typically the rectified linear unit (ReLU) for hidden layers in modern applications [14].

MLPs are trained using the backpropagation algorithm, which minimizes a loss function via gradient descent. For regression tasks, the Mean Squared Error (MSE) is commonly used and is defined as:

$M S E=\frac{1}{n} \sum_{i-1}^n\left(y_i-\hat{y}_i\right)^2$                               (4)

where, $y_i$ is the true illuminance value, $\widehat{y_{\imath}}$ is the predicted value, and n is the total number of samples.

In this study, the MLP consisted of an input layer with four normalized features (R_avg, G_avg, B_avg, GS_avg), two hidden layers (200 and 100 neurons) with ReLU activation, and a single output neuron predicting illuminance (lux). Adam optimizer was used for training due to its adaptive learning rate and efficiency on noisy data [15]. Figure 5 illustrates the architecture, where features are transformed to learn high-level patterns.

505.png

Figure 5. MLP architecture for illuminance prediction

The final configuration of the MLP model was obtained through grid search hyperparameter tuning. The optimal parameters are listed in Table 1, and reflect the settings that yielded the best validation performance using k-fold cross-validation.

These hyperparameters were selected based on systematic experimentation and validation results. The MLP Regressor, being a type of feedforward neural network, is capable of learning complex nonlinear relationships between input features and target outputs. Its flexibility makes it suitable for a wide range of problems. However, MLPs are sensitive to the choice of hyperparameters, require more computational resources for training, and may suffer from overfitting with small datasets.

Table 1. The parameters of the MLP model

2.2.2 Random Forest Regressor

The Random Forest Regressor (RFR) is an ensemble learning method that builds multiple decision trees and averages their outputs, making it effective for nonlinear regression. It enhances generalization and reduces overfitting through bootstrap sampling and random feature selection during training [16]. Each decision tree is trained on a different bootstrap sample (i.e., random sampling with replacement) of the training data. During the construction of each tree, only a random subset of features is considered at each split [17], encouraging decorrelation between trees and enhancing the robustness of the ensemble. The overall prediction of the Random Forest is computed as the average of the predictions made by the individual trees:

$\hat{y}=\frac{1}{T} \sum_{t=1}^T f_t(x)$                                (5)

where, $\hat{y}$ is the final prediction, $T$ is the number of decision trees, and $f_t(x)$ is the prediction made by the t -th tree made by for input [18].

Each tree predicts by recursively splitting data based on features to minimize node impurity. For regression, the impurity measure is typically the Mean Squared Error (MSE), defined for node m as:

$M S E_m=\frac{1}{N_M} \sum_{i \in N_m}\left(y_i-\bar{y}\right)^2$                               (6)

where, $N_m$ is the set of samples reaching node $m, y_i$ is the target value of sample $i, y_{m}$ is the average target value of the samples in node $m$ [19].

In this study, the Random Forest model was optimized using grid search and k-fold cross-validation, which result the best structure summarized in Table 2.

Table 2. The parameters of the RFR model

This configuration allowed the model to learn complex patterns present in the dataset while avoiding overfitting. The number of trees was set to 200 to provide a robust ensemble and to avoid a high increase in training time. Minimum sample constraints used to ensure statistical significance at each node, and by using all four normalized input features in model training allowed the model to take advantage of both color and luminance cues in estimating illuminance.

2.2.3 XGBoost Regressor

The eXtreme Gradient Boosting (XGBoost) Regressor is an advanced gradient boosting algorithm which is characterized by an ameliorated accuracy. It builds an ensemble of decision trees sequentially where each tree tries to correct the errors made by the previous tree by minimizing a loss function via gradient descent this prediction process must sequentially traverse multiple decision trees. Even during inference, this sequential evaluation of hundreds of trees can result in slower prediction speeds compared to simpler models. The model predicts output as a sum of functions, each corresponding a decision tree:

$\hat{y}_i=\sum_{K=1}^K f_K\left(x_i\right), f_K \in \Gamma$                         (7)

where, $\Gamma$ is the space of regression trees, $x_i$ is the feature vector for instance $K$ is the number of trees. The optimization objective consists of a loss function measuring prediction error and a regularization term to penalize model complexity:

$L(\Phi)=\sum_{i=1}^n l\left(y_i, \hat{y}_i\right)+\sum_{k=1}^k \Omega\left(f_k\right)$                          (8)

with $l$ as the loss function, $\Omega(f)$ as the regulation term where $T$ is the number of leaves in the tree, $w$ are the leaf weights, and $\gamma, \lambda$ regularization parameters, this formulation enables XGBoost to generalize well on unseen data by controlling overfitting [20].

In this study, XGBoost was trained using the normalized features R_avg, G_avg, B_avg, and GS_avg as input, with illuminance (ILS) as the output. Grid search and cross-validation were used to identify the optimal hyperparameters, as shown in Table 3.

Table 3. The parameters of the XGboost estimator

Every tree in the ensemble, splits the input space into distinct subsets depending on the most informative input attributes. As trees are added sequentially, the model corrects its previous errors iteratively, leading to a robust predictor with the ability to effectively estimate illuminance for different lighting conditions.

2.2.4 Training and validation

In this study, a well-structured and systematic strategy was applied to carry out the training and validation of the three regression models in order to identify the most effective one for practical deployment in real-world lighting assessment tasks. The primary objective was to construct machine learning models capable of predicting illuminance values from features extracted from smartphone-captured indoor images with both high accuracy and strong generalization performance.

Each one of the three models was trained on a preprocessed and normalized dataset comprising statistical features extracted from 2×2 image patches and their corresponding illuminance values (in lux). Training began by randomly splitting the dataset into 80% training and 20% validation subsets, ensuring that models were evaluated on unseen data. To ensure a fair and consistent basis for comparative analysis all models utilized the same training-validation split.

Subsequently, each model underwent hyperparameter tuning via Grid Search Cross-Validation to identify the combination of parameters yielding the optimal performance on the validation set. The tuned parameters for the MLP Regressor model, the tuned are the hidden layer sizes, activation function, the solver, the alpha value, and the learning rate. For the RFR model, the tuned parameters were number of estimators, maximum depth, minimum samples split, and minimum samples leaf. Finally, for XGBoost Regressor the hyperparameters are the number of trees, the maximum tree depth, learning rate, loss function, and the regularization parameters.

To optimize the models hyperparameter Grid Search with k-fold Cross-Validation was performed this process is a robust technique for evaluating parameter combinations and avoiding overfitting. The following hyperparameters were tuned for each model:

·MLP Regressor: number of hidden layers and their sizes, activation function, solver (optimizer), L2, batch size, loss function, and learning rate.

·Random Forest Regressor: number of trees (n_estimators), maximum tree depth, minimum samples required to split a node, and minimum samples per leaf.

·XGBoost Regressor: number of trees, maximum tree depth, learning rate, loss function, and regularization parameter L2.

The process of evaluating the regression models requires robust and interpretable evaluation metrics. In this study we employed three widely accepted indicators to assess performance:

$M A E=\frac{1}{n} \sum_{i=1}^n\left|y_i-\hat{y}_i\right|$                              (10)

MAE’s reliance on absolute values makes it less sensitive to outliers than squared-error measures like Mean Squared Error (MSE). In addition to MAE, Root Mean Squared Error (RMSE) was also used, which is defined as:

$R M S E=\sqrt{\frac{1}{n} \sum_{i=1}^n\left(y_i-\hat{y}_i\right)^2}$                             (11)

RMSE retains MSE’s sensitivity to large errors while expressing them in the target variable’s units (lux), enhancing interpretability. It effectively highlights significant prediction errors, which are crucial in lighting-sensitive settings like laboratories.

To measure the proportion of variance in the target variable that can be explained by the model, the R² score was used:

$R^2=1-\frac{\sum_{i=1}^n\left(y_i-\hat{y}_i\right)^2}{\sum_{i=1}^n\left(y_i-\bar{y}\right)^2}$                          (12)

where, $\bar{y}$ is the mean of the true illuminance values. A perfect model yields $R^2=1$, whereas an $R^2$ close to 0 indicates poor model performance. $R^2$ provides an intuitive measure of model fit and is useful for comparing model generalization on validation data.

Figure 6 illustrates the training and validation pipeline employed in this study, depicting the sequential process for developing and evaluating the illuminance estimation models. 

506.png

Figure 6. Training and validation process

The dataset was first split into training and test sets. Using cross-validation on the training data, hyperparameters for the MLP, Random Forest, and XGBoost Regressors were tuned to optimize model fit and generalization.

The best models were then retrained on the full training set and evaluated on the unseen test set. Performance was assessed using MAE, RMSE, and R² to measure accuracy and generalization.

2.3 Mobile app functionality

The lightweight mobile application developed in this study is able to make a real-time illuminance analysis using a trained machine learning model to process camera input, generate an illuminance heatmap, and provide actionable lighting recommendation entirely on-device, without relying on external computation or connectivity.

As shown in Algorithm 1, the app captures an image, divides it into 2×2 patches, and extracts average R, G, B, channels and grayscale intensity from each patch. These normalized inputs are then passed into a TensorFlow.js-based model deployed within the mobile environment, enabling real-time inference for each patch.

The results form a color-coded heatmap overlay on the original image, while the average illuminance is compared against standards for the selected indoor environment.

To determine whether the measured value falls outside or within the acceptable range recommended by CIE and IES associations, Table 4 illustrates the recommended illuminance levels for most possible indoor settings.

Table 4. Recommended lux level

The app automatically provides lighting adjustment recommendations based on the computed average illuminance. When the lighting level falls outside the recommended range, a “Highlight Poorly Illuminated Areas” button appears.

Upon activation, the overlay updates to visually mark over-illuminated regions in red and under-illuminated areas in blue, helping users easily identify specific lighting issues.

The user interface flow of the mobile application is illustrated in Figure 7 and consists of three main screens. The first screen allows users select an indoor environment (e.g., living room, classroom, retail), each of which is associated to a recommended illuminance range.

507.png

Figure 7. User interface flow of the mobile application

After selection, the app moves to the second screen, which streams live video from the smartphone’s rear camera.

Pressing the “Scene Analysis” button freezes the current frame and transitions to the third screen, displaying the analysis overlay and modal recommendations This streamlined flow from environment selection to live preview and finally to interactive feedback ensures an intuitive and efficient user experience.

This study introduced a low-cost, smartphone-based tool capable of real-time indoor illuminance estimation and mapping, leveraging deep learning and classical machine learning models to support compliance with lighting design standards.

The proposed mobile application enables users to assess lighting conditions visually and numerically, eliminating the need for costly lux meters or specialized instrumentation. Through a platformatic development pipeline encompassing data collection, model training, validation, and integration into a real-world application, we demonstrated that compact devices can provide accurate and spatially-resolved illuminance feedback.

Experimental results showed that among the three tested models Multi-Layer Perceptron, Random Forest Regressor, and Gradient Boosting Regressor; the Random Forest Regressor demonstrated the best trade-off between prediction accuracy (MAE: 21.25, R²: 0.97) and inference speed (≈ 40 ms). Furthermore, Mean Absolute Error Percentage (MAEP) of 2.43% across multiple desk locations in real-world environments, validated the consistency and spatial reliability of the predicted illuminance distribution under varied conditions. The complete application maintained a total processing time of under one second, rendering it suitable for responsive mobile-based lighting analysis. By offering real-time heatmap visualizations and context-aware lighting recommendations within a single portable platform, this application presents new opportunities for intuitive lighting diagnostics in residential, educational, healthcare, and commercial spaces. The user-friendly interface and compatibility with consumer-grade smartphones further enhance its accessibility and scalability.

However, it is important to acknowledge that the platform was only evaluated in a limited range of indoor environments using a single smartphone model. Broader deployment scenarios involving diverse room geometries, surface reflectance, and device-specific camera characteristics should be explored to confirm the model’s robustness and generalizability.

Future work will focus on addressing the current limitations and enhancing the platform's robustness, scalability, and adaptability. A key priority will be the evaluation of system performance across a broader range of real-world indoor environments, including residential, industrial, and commercial settings, each with distinct lighting configurations, surface textures, and spatial geometries. This expanded validation will help assess the model’s generalization capabilities beyond controlled office and classroom scenarios.

To improve cross-device consistency, future versions may incorporate device-specific calibration routines or normalization layers to account for variability in camera hardware and built-in image processing algorithms. In parallel, advanced preprocessing techniques such as high-dynamic-range (HDR) fusion or exposure correction could be explored to better handle scenes with complex or uneven illumination patterns, including glare, shadows, and mixed lighting sources.

From a computational standpoint, optimizing the model for real-time inference on low-end and mid-range smartphones will be essential. This may involve quantization, model pruning, or knowledge distillation to reduce memory and processing demands without compromising accuracy. Furthermore, the integration of lightweight edge computing frameworks could ensure smooth performance while preserving offline functionality.

Finally, future iterations of the system could incorporate temporal illumination tracking and user feedback mechanisms. These enhancements would enable the platform to learn from environmental patterns and user behavior over time, ultimately supporting intelligent daylight harvesting strategies and adaptive lighting control systems that respond dynamically to both spatial and temporal context.