Robust Deep Learning Based Fruit Recognition System for Autonomous Harvesting System in Complex Cashew Orchards

Sudha C, K. JaganMohan,
Robust Deep Learning Based Fruit Recognition System for Autonomous Harvesting System in Complex Cashew Orchards,
Smart Agricultural Technology,
2025,
101769,
ISSN 2772-3755,
https://doi.org/10.1016/j.atech.2025.101769.
(https://www.sciencedirect.com/science/article/pii/S2772375525010007)
Abstract: Deep Learning algorithms have recently exhibited enhanced efficiency in addressing challenging complex computer vision problems particularly in agricultural automation. Cashew fruit identification in complex orchard environments is a crucial task for intelligent orchard management, robotic harvesting, and automated yield estimation in precision agriculture. To accomplish swift and accurate Cashew fruit identification amidst intricate environmental conditions, this research suggests a fruit detection technique utilizing Deep Learning algorithms. This research introduces an effective algorithm for Cashew fruit detection built on the Single Shot MultiBox Detector (SSD) with a MobileNetV2 backbone, enhanced by a streamlined Feature Pyramid Network (FPN-Lite). This combination results in a deep neural network that is both computationally efficient and high-performing, making it ideal for use in mobile and embedded vision settings. To minimize computational load while preserving feature representation, MobileNetV2 utilizes depth wise separable convolutions along with inverted residual blocks. Meanwhile, the FPN-Lite component allows for the integration of multi-scale features, which enhances detection accuracy for objects of different sizes. The proposed SSD MobileNetV2-FPNLite design is especially effective for real-time detection in agriculture, even under resource-constrained conditions, because the SSD architecture supports comprehensive object detection by enabling simultaneous classification and localisation within a single forward pass. Under a variety of difficult orchard conditions, the suggested model was trained to correctly identify cashew apples at three phases of maturity: immature, unripe, and ripe. With an average inference time of 8 seconds per high-resolution image, the system attained a mean detection accuracy of 92.26%. The model successfully identifies a variety of cashew fruit colours, such as red, yellow, and mixtures of these colours. Additionally, the system quickly and accurately recognises various phases of maturity with a high recognition rate and can identify individual cashew fruit within clusters, which helps the harvesting robot with its ongoing fruit plucking process and makes yield estimation easier. In general, this paper tackles several important issues related to fruit identification and localisation in intricate orchard settings.
Keywords: Autonomous harvesting system; Cashew Fruit Detection; Deep Learning; SSD MobileNetV2FPNLite; YOLOv4; Faster RCNN; Robotic Harvesting; Agricultural Automation; Neural network