CamTran: An Intelligent Automatic System for the Pest Infestation Assessment

In the vast expanse of the tea industry, addressing pest infestations on tealeaves and improving vegetation indices has long been a challenge. While these issues may seem minor on the surface, in reality, they can lead to substantial financial strain. Beyond the immediate, monetary implications, the ripple effect of these challenges can profoundly impact local economies, disrupt trade, and jeopardize the livelihoods of those deeply entrenched in the world of tea cultivation. Convolutional Neural Networks (CNNs) are widely acknowledged for their capability to accurately identify and classify complex characteristics in images, especially those depicting disease or pest-afflicted tea leaves. The technology’s premise lies in autonomously extracting features from imagery, allowing for accurate and timely identification, which can potentially lead to early interventions. The process of capturing images of tealeaves in their natural habitat is fraught with complexities. Factors like fluctuating lighting conditions, owing to the time of day or changing weather patterns, can significantly alter an image’s clarity. Cultivating tea in the plantation, it has been observed that the holistic life cycle stages of several pests on tea leaves have been detrimentally affecting both the yield and quality of the tea leaves, demanding a comprehensive automated assessment strategy. Recognizing these shortcomings, the paper adopts a fundamental shift in approach and implements CamTran. The architecture, rooted in the Vision Transformer model, adopts a specialized attention mechanism to discern and prioritize salient features within localized image segments, ensuring efficient and relevant feature extraction. The Non-local Blocks, on the other hand, serve a strategic purpose by capturing and integrating long-range dependencies and spatial relationships in the image, providing a holistic understanding of the context. Additionally, the Atrous Spatial Pyramid Pooling (ASPP) module is integrated to gather information on objects at multiple scales, while not escalating the count of parameters or computational load. Depth wise Separable Convolutions, a prominent feature, enhance the network’s ability to recognize and discern between nuanced patterns by reducing computational cost and maintaining performance. Regarding optimization, a blend of stochastic gradient descent with momentum and adaptive learning rates was utilized. Recognizing the potential for over fitting, especially given the complexity of the architecture, L2 regularization coupled with dropout techniques was implemented. An extensive dataset, methodically divided with 60% allocated for training, 30% for validation, and a dedicated 10% exclusively for testing purposes. The engine consistently displayed accuracy levels hovering between 94-96%. For a more holistic evaluation, CamTran was juxtaposed against established benchmarks like the Microsoft COCO metrics, registering an impressive 96% in Average Precision (AP) at 0.5 and maintaining commendable scores across the other evaluation parameters, including Recall, F1-score, and loss metrics.