Dual-Stream MobileNetV3Small Fusion Architecture for Enhanced Multi-Class Detection of Arecanut Diseases Using Hybrid Feature Embedding

Dinesh S, P. Sridhara Acharya

Abstract: This research proposes an alternative approach for diagnosing multiple arecanut diseases by introducing a hybrid architecture known as MobileNetV3Small-CM FusionNet. Unlike conventional image-only models, the framework merges two complementary information streams: compact deep representations derived from the MobileNetV3Small backbone and statistical color descriptors computed through color moments. By integrating these feature types, the model is able to capture subtle chromatic variations and disease-specific patterns that are often difficult for lightweight CNNs to distinguish. The dataset used in this study comprises labeled images of both healthy and infected arecanut samples and is partitioned for training, validation, and testing. To assess effectiveness, the proposed system is compared with a standard CNN and the original MobileNetV3Small model. Evaluation results indicate a notable performance advantage for the fusion-based approach, which records a test accuracy of 99.54% and consistently high precision, recall, and F1-scores across all nine categories. In comparison, the benchmark models achieve considerably lower accuracy levels of 93.74% and 83.87%. These findings highlight the value of combining handcrafted statistical cues with deep feature embeddings for reliable and robust plant disease identification.

Keywords: MobileNetV3Small, color moments, hybrid fusion model, arecanut disease detection, plant image classification

How to Cite?: Dinesh S, P. Sridhara Acharya, "Dual-Stream MobileNetV3Small Fusion Architecture for Enhanced Multi-Class Detection of Arecanut Diseases Using Hybrid Feature Embedding", Volume 14 Issue 11, November 2025, International Journal of Science and Research (IJSR), Pages: 1737-1741, https://www.ijsr.net/getabstract.php?paperid=SR251124133511, DOI: https://dx.doi.org/10.21275/SR251124133511