DEVELOPING AN ANDROID APPLICATION FOR RECOGNIZING SOUTHERN VIETNAMESE VEGETABLES

Abstract

In this thesis, we demonstrate the method of building the Android application, covering the dataset used, the training model, and the experimental results in Southern Vietnamese vegetables recognition. We conducted experiments with state-of-the-art computer vision models, namely YOLOv10 and YOLOv11, alongside our proposed lightweight and accurate model, YOLOv11-GhostNet. Our approach optimizes YOLOv11 by replacing some of the layers in the backbone with GhostNet module, significantly reducing model size, and incorporating the Convolutional Block Attention Module (CBAM) to enhance feature extraction. The proposed method is evaluated using ViSVeggie dataset contains 26 distinct classes. We evaluate the performance of object detection using Mean Average Precision (mAP) metric across varying IoU thresholds (mAP50-95). The experimental results show that our model achieves an mAP50-95 value of 0.793, which is 3.3% higher than YOLOv10 (0.760) and 2.1% higher than YOLOv11 (0.772), fewer parameters and reducing computation load compared to its counterparts, while maintaining a model size of 16.5MB and inference time of 6.0ms.