Introduction

The identification and quality control of Ayurvedic herbs are critical to ensuring the safety, efficacy, and consistency of traditional medicinal formulations. Ayurveda, an ancient system of medicine, relies heavily on the use of herbs, including fruits and seeds, to treat a variety of ailments. However, the accurate identification of dry herbs, especially in their dry forms, poses a significant challenge due to the morphological similarities among species, the presence of adulterants, and the lack of standardized methods for quality control. And hence adulteration is prevalent; for example, seeds from Carica papaya are sometimes misbranded as Piper longum, and Cinnamomum cassia is often sold as Cinnamomum zeylanicum, despite significant differences between them.^[1] The Ayurvedic Pharmacopoeia of India (API)^[2] prioritizes macroscopic examination based on visual characteristics for identifying herbal drugs. Acc. to WHO macroscopic examination of herbal materials provide a straightforward and rapid method to determine their identity.^[3-5] However, traditional methods face challenges in a global market, necessitating advanced technologies such as Computer Vision Technology (CVT) and molecular analysis. Applications like Ayurleaf^[6] and Leafsnap^[7] are example of this. Similarly, Praman AI,^[8] an app focused on the cardamom trade, uses AI for post-harvest quality assessment, streamlining agricultural processes and ensuring product quality. Computer vision and Artificial intelligence-based modelling are explored by many researchers in recent time for industrial application. Despite the success of these technologies, the application of computer vision for dry Ayurvedic herb.

The emphasis of presented study was on the training a model on the macroscopic features of 50 (fruits and seeds) commonly traded dry herbs through Computer Vision technology using the Multiclass classification and Convolutional neural network approach. The basic concept of NN and Artificial Intelligence with the proposed architecture with detailed parameters are given in the manuscript. The results have been explored and discussed along with the conclusion of paper with further future scope of the presented study. Here is the list of Dry Ayurvedic herbs which were used in the study.

Neural Network

An Artificial Neural Network (ANN) is designed to replicate the structure and functionality of neurons in the human brain, enabling it to understand non-linear physical processes.^[9,10] ANNs are built by creating artificial neurons and interconnecting them in a way that emulates the behaviour of interconnected brain cells. When these networks contain a large number of neurons (nodes) arranged in layers, they are referred to as Deep Neural Networks (DNNs). NNs have revolutionized artificial intelligence by providing the capability to understand and solve a wide range of complex real-world problems. These networks are characterized by an architecture that includes a single input layer, multiple hidden layers for learning non-linear patterns from input data, and a final output layer. Each layer contains numerous neurons, where each neuron processes multiple inputs according to an activation function and transmits the computed output to the subsequent layer, as illustrated in Fig.1 The inputs (“x1, x2, … xi … xn”) are referred to as input nodes, while the outputs (“y1, y2, … yi … ym”) are called output nodes of the network.^[11]

For a logistic regression problem, there is typically one output node; for binary classification, number of output nodes is two; and for more complex problems, the number increases. A typical DNN, depicted in Fig.2, comprises multiple hidden layers that enable network to learn complex features and non-linear patterns from input data. Each layer processes outputs of the previous layer using its neurons (nodes), and final layer produces the overall output of the network. The layers in a DNN are connected by a set of weights (“w1, w2, … wi … wl”), which are adjusted during training process to optimize performance.^[12,13]

Figure 1: Basic structure of Neural Network

For image analysis, a powerful computer system featuring Intel Xeon processors and NVIDIA Tesla V100 GPUs was utilized. Additionally, various software tools, including Photoshop, TensorFlow, Keras, and Python libraries like OpenCV were employed to process and analyze the images effectively.

Methodology

1. Data Collection: Images were collected to cover diverse views for improved model
2. Pre-processing and Augmentation: Images were resized, normalized, and augmented for better model generalization.
3. CNN Model Development: A convolutional neural network (CNN) was trained for herb
4. API and UI Development: An API and user-friendly interface enabled image uploads and classification results.
5. System Integration and Testing: The system was integrated, tested, and deployed, followed by real-world evaluations.

Proposed Work

Ayurvedic herb identification through computer vision has been addressed using a Convolutional Neural Network (CNN). Traditional identification methods based on macroscopic examination are complex and time taking so need experts, making AI-driven approaches viable alternatives. The dataset comprising high-resolution images of dry Ayurvedic herbs were pre-processed and used to train the CNN model. Input parameters such as image size, batch size, and data augmentation techniques (random rotations, flips, and zooms) were optimized to enhance model generalization.

Figure 4: Accuracy Over epochs

Conclusion

The CNN model developed for the identification of dry Ayurvedic herbs (fruits and seeds) demonstrated promising potential, achieving a peak training accuracy of 91.86% and validation accuracy ranging between 81% and 83%. This performance highlights the model's capability to extract meaningful features for herb classification. However, misclassifications were observed, particularly among visually similar herbs, indicating areas for improvement. The architecture's inference time of 36 milliseconds per step and robust performance on distinct herb species suggest its practical applicability for real-time herb identification. Despite its effectiveness, the model's generalization issues underline the need for further refinement. To enhance the model's robustness and scalability, several recommendations are proposed. Expanding the dataset to include diverse species, plant parts, and variations in drying and processing conditions, alongside capturing images from multiple angles and lighting conditions, could improve generalization. Addressing class imbalances through augmentation or resampling strategies and optimizing hyperparameters like learning rate, batch size, and dropout rate could mitigate overfitting. Additionally, incorporating advanced regularization techniques, such as L2 regularization or early stopping, and leveraging enhanced computing systems, including GPUs or TPUs, could further refine the model's performance. Implementing cloud-based platforms for distributed training and applying efficient pre-processing techniques like noise reduction and contrast enhancement would also contribute to more accurate classifications. These enhancements would establish a reliable framework for quality assurance in herbal medicine, fostering advancements in the identification and utilization of Ayurvedic herbs.

References