Visionary Banknotes Denomination Recognition: Empowering the Visually Impaired with Deep SqueezeNet Model

Identifying currency is a major challenge for individuals with visual impairments. With millions of people affected by blindness worldwide, there is a strong need for assistive technologies that can recognize currency [1]. Existing methods fall into two categories: scanner-based and camera-based systems. Scanner-based systems require the full image of a banknote for recognition, whereas camera-based systems can identify currency from just a portion of the image. Our approach uses a camera-based system that recognizes Egyptian banknotes even when only a part of the image is captured.

Deep learning plays a key role in image recognition. It automatically extracts important features from images using convolutional layers, which are effective in learning complex patterns. However, deep learning models often need extensive datasets for effective training [2]. To address this issue, we use transfer learning, which leverages knowledge from pre-trained models. This method reduces both the need for large datasets and training times [3].

Our method depends on SqueezeNet, a compact pre-trained deep learning model, to derive visual features from images [4]. Once extracted, these features are classified using and Support Vector Machines (SVM) classifier. We created a dataset of Egyptian currency that includes notes of one, five, ten, twenty, fifty, one hundred, and two hundred pounds. The images include various orientations and partial views of banknotes, making the system adaptable and effective in real-world conditions.

The remainder of this paper is structured as follows. Section 2 provides an overview of the related research. Sections 3 and 4 cover the fundamentals of transfer learning and the details of the SqueezeNet architecture. Section 5 outlines the application of SVM for classification. Section 6 describes the proposed framework, and Section 7 focuses on its implementation. Finally, the last section summarizes the main conclusions and insights drawn from this study.

Convolutional Neural Networks (CNNs) have become a cornerstone of image-based research, as widely demonstrated in the literature [9, 10]. These models excel at extracting complex low-level image features. Yet, training deep CNNs typically requires large datasets. To overcome this limitation, researchers often use transfer learning, based on models pre-trained with extensive data collections like ImageNet, which are adapted for new tasks. These pre-trained models can recognize general features such as edges and contours that are common across various types of images, making them transferable to different domains [11-14] (see Figure 1).

Figure 1. Traditional machine learning in comparison to transfer learning

Figure 2. SqueezeNet architecture

Transfer learning thus serves as a bridge, allowing knowledge from large-scale datasets to be applied to smaller, task-specific datasets. This not only reduces the need for extensive training data but also increases model efficiency and expands the applicability of deep learning across diverse use cases [3].

In this study, we utilize the CNN architecture known as SqueezeNet [15] (see Figure 2). SqueezeNet is specifically designed to offer high performance, while significantly reducing model size, computational cost, and inference time. Its compact architecture enables efficient feature extraction with minimal resource consumption, making it suitable for deployment on a wide range of devices and platforms. By balancing accuracy and efficiency, SqueezeNet supports practical deep learning applications, especially when computing resources are limited [16].

The SVM classifier is a prominent machine learning technique employed for the classification of sample data, leveraging input features acquired during the learning process [17]. The SVM is a highly valuable tool in discerning the optimal differentiating hyperplane among distinct classes within the target feature space.

One of SVM classifier fundamental strengths lies in its ability to discern intricate patterns within data, making it particularly well-suited for tasks involving binary classification. By meticulously identifying the most suitable hyperplane, SVM classifier optimally separates data points representing diverse classes. This powerful capability is at the core of the SVM classifier effectiveness in solving classification problems across various domains (see Figure 5).

Figure 5. Support Vector Machines classifier concept

The primary goal here is to determine a hyperplane that possesses the maximum margin, which refers to the greatest separation distance between data points representing different classes. Maximizing this margin distance holds significant importance as it imparts robustness to the classification model, enhancing its ability to classify future data points with greater confidence and precision [18].

By striving to create a broader margin, we essentially carve out a more substantial and well-defined boundary between classes. This not only aids in the accurate classification of existing data but also fortifies the model resilience in handling novel data points, ensuring that the classification decisions made are more reliable and effective [19].

One of the significant advantages derived from the utilization of SqueezeNet lies in its remarkable efficiency in handling smaller datasets. Leveraging a pre-trained network, the system adeptly extracts features from the training images, with the dataset thoughtfully partitioned into 70% of the images for training, and the last 30% dedicated for testing.

In the training phase, the SVM classifier is effectively trained using the extracted features. During the testing phase, SqueezeNet continues to play a pivotal role, extracting features from the test images and subsequently forwarding them to the SVM classifier for accurate labeling, a process vividly represented in Figure 7 for clarity. To measure the efficacy pertaining to the proposed framework, several metrics have been used, namely accuracy, recall, and F1-score [21].

Accuracy reflects the overall accuracy by comparing the sum of true positives and true negatives against the total predictions [21].

Accracy $=\frac{T_p+T_N}{T_p+T_N+F_p}$

where,

$T_p$: True Positives (correctly predicted positive class)

$T_N$: True Negatives (correctly predicted negative class)

$F_p$: False Positives (incorrectly predicted positive class)

$F_N$: False Negatives (incorrectly predicted negative class)

Recall measures how many actual positive instances were correctly predicted [21].

Recall $=\frac{T_P}{T_P+F_N}$

F1-Score is a metric that balances precision and recall through their harmonic mean, and is particularly useful in imbalanced datasets [21].

$\mathrm{F} 1-$ Score $=\frac{\text { Precision } \cdot \text { Recall }}{\text { Precision }+ \text { Recal }}$

Confusion matrix is a performance measurement tool used to evaluate the results of a classification model. It shows how often actual classes are correctly and incorrectly predicted.

The culmination of this meticulous process is revealed through simulation results, which unequivocally affirm the effectiveness of the proposed system. The system demonstrates an impressive recognition accuracy of 99% without augmentation and 99.5% with augmentation, as prominently depicted in Figures 8, and 9, respectively. These results underscore the system capacity to accurately identify Egyptian banknotes, reinforcing its practicality and utility in real-world applications.

Figure 8. Confusion matrix of the proposed system (SqueezeNet+SVM classifier) without augmentation

Figure 9. Confusion matrix of the proposed system (SqueezeNet+SVM classifier) with augmentation

Figure 10. Samples of tested images with folding, blurring, reduced resolution and salt & pepper noise

Furthermore, a selection of the sample images that underwent testing is showcased in Figure 10. This visual representation offers a practical glimpse into the real-world applicability of the system and how it accurately performs in the context of recognizing Egyptian banknotes. These sample images serve to illustrate the system precision and its ability to effectively classify and identify banknote denominations.

The experimental findings confirm the efficiency of the proposed system in accurately recognizing banknotes, even under challenging conditions such as partial occlusion, folding, blurring, or reduced image resolution and in the case of images with salt & pepper noise as in Figure 10. The system consistently achieves high recognition accuracy, demonstrating notable robustness and adaptability in practical, real-world environments.

The results in Table 2 show that data augmentation has a positive impact pertaining to the results of the proposed system. Without augmentation, the SqueezeNet+SVM model achieved an already high accuracy of 99%. However, with data augmentation, the accuracy is further improved to 99.5%, indicating enhanced generalization and robustness of the model when exposed to variations in the training data. This demonstrates that even small improvements in data diversity can contribute to more reliable classification, especially in real-world scenarios where input images may vary in quality or conditions.

Table 2. Recognition accuracy of the proposed method with and without augmentation

Table 3. Assessment of the proposed system relative to other different methods

To evaluate the performance of the proposed system, it was compared with several other classification algorithms, namely the Multi-Layer Perceptron (MLP) [22] and Random Forest classifier [23]. The same SqueezeNet model was used for feature extraction in all cases to ensure a fair comparison. The proposed system (SqueezeNet+SVM) achieves high scores in accuracy, recall, and precision, in addition to maintaining an acceptable inference time (see Figures 11-13 and Table 3).

Figure 11. Confusion matrix of the SqueezeNet+Random Forest method

Figure 12. Confusion matrix of the SqueezeNet+Multi-Layer Perceptron (MLP) method

Figure 13. Confusion matrix of the MobileNetv2 + Support Vector Machines (SVM) method

The comparative results in Table 2 clearly demonstrate that the SqueezeNet+SVM combination (the proposed system) outperforms both the Random Forest and MLP classifiers in terms of accuracy, precision, and recall, achieving an impressive 99.5% accuracy with high consistency across metrics. While Random Forest classifier performs well with a strong accuracy of 97.4%, it requires a longer inference time (7.4 seconds), indicating potential inefficiencies in real-time applications. On the other hand, the MLP classifier offers the fastest inference (~0.02 seconds), but lags in recognition performance, especially with a recall of 92.8%, showing that it may struggle with generalization. The MobileNetv2 gives an accuracy of 97.4%, and it requires a time of 0.91 seconds as in Figure 13.

The proposed SqueezeNet+SVM model provides the best balance, combining exceptional classification performance with a reasonable inference time (~0.87 seconds). This demonstrates its robustness and suitability for real-time or near-real-time banknote recognition tasks, especially under non-ideal conditions such as partial visibility or image degradation and images with Salt & Pepper noise.

Also, a comparative analysis was conducted, pitting the results obtained against those of CNN models featuring two, three, and four layers. A summary of these comparative results is tabulated in Table 3. This comparative analysis underscores the superior performance of the proposed system in terms of banknote recognition accuracy, especially under challenging conditions, showcasing its practicality and versatility (see Figures 14-16).

While the proposed currency recognition system achieves high accuracy and robustness in classification, its practical deployment requires careful consideration of computational efficiency and hardware constraints. In particular, implementing the system on embedded platforms or mobile devices necessitates optimization of model size, memory footprint, and inference speed. The use of lightweight architectures such as SqueezeNet already contributes to lower computational demands (model size ~5 MB), making the model suitable for resource-constrained environments. Preliminary experiments indicate an average inference time of approximately 0.87 seconds on a standard desktop CPU, showing potential for real-time or near-real-time processing with further optimization.

Figure 14. Two-layer Convolutional Neural Network results

Figure 15. Three-layer Convolutional Neural Network results

Figure 16. Four-layer Convolutional Neural Network results

To enhance compatibility with mobile and embedded devices, model compression techniques such as quantization, pruning, and conversion to formats like TensorFlow Lite or ONNX can be applied. These techniques allow for reduced model size and faster execution without significant loss in accuracy. Potential deployment targets include Android smartphones, which offer high processing capabilities, integrated cameras, and support for mobile Artificial Intelligence (AI) frameworks (e.g., TensorFlow Lite, ONNX Mobile). Alternatively, platforms such as the Raspberry Pi or NVIDIA Jetson Nano could serve as cost-effective embedded options, though they may require GPU acceleration or additional optimization to meet real-time constraints.

Future work will explore these deployments, ensuring the system usability in real-world settings, where fast and reliable currency recognition is essential.

Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R138), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.