A Novel Approach for Efficient Detection of Lotus Seedpod Maturity Using Compressed Models

This study aims to achieve efficient and accurate detection of lotus seedpod maturity in complex environments. To address the challenges associated with existing object detection algorithms, which often involve numerous model parameters and high computational loads that hinder deployment on resource-limited mobile terminals, we propose an efficient lotus seedpod maturity detection method based on a compressed model. First, data augmentation techniques were employed to enhance the diversity of lotus seedpod image samples. Next, a model was established under the CSPDarknet53 framework, incorporating a fast spatial pyramid pooling module to detect the maturity of lotus seedpods. To simplify the model and improve detection speed, a channel pruning algorithm was utilized for model compression. Finally, the compressed model was fine-tuned to restore accuracy. Experimental results demonstrate that the compressed model reduced the number of parameters, model size, and inference time by 72.96%, 70.96%, and 26.92%, respectively, achieving a mean Average Precision (mAP) of 99.2%, representing only a 0.20% decrease compared to the original model. Compared to Faster R-CNN, SSD, YOLOv5, YOLOv7-tiny, YOLOv10 and YOLOv11 models, the proposed model significantly reduces parameter count, computational load, and model size while maintaining a high mAP, thus demonstrating feasibility for rapid and accurate detection of lotus seedpod maturity. Additionally, a prototype testing platform for lotus seedpod maturity detection was established, and the compressed model was deployed on an NVIDIA® Jetson Xavier NX mobile terminal. The testing results indicated that the compressed model size was 3.95 MB, with a detection speed of 416.67 frames per second, meeting real-time detection requirements and confirming its applicability for low-computational-capacity mobile terminals. This research provides valuable technical support for the subsequent development of lotus seedpod harvesting robots.