摘要
Recognizing fruit and vegetable is crucial for improving processing efficiency, automating harvesting, and facilitating dietary nutrition management. The diverse applications of fruit and vegetable recognition require deployment on end devices with limited resources, such as memory and computing power. The key challenge lies in designing lightweight recognition algorithms. However, current lightweight methods still rely on simple CNN-based networks, which fail to deeply explore and specifically analyze the unique features of fruit and vegetable images, resulting in unsatisfactory recognition performance. To address this challenge, we propose a novel lightweight recognition network termed Channel Grouping Vision Transformer (CGViT). CGViT utilizes a channel grouping mechanism and half-convolution to enhance feature extraction capability while reducing complexity. This design enables the model to capture three discriminative types of features from images. Subsequently, the Transformer is employed for feature fusion and global information extraction, ultimately creating an efficient neural network model for fruit and vegetable recognition. The proposed CGViT approach achieved recognition accuracies of 71.26%, 99.99%, 98.92%, and 61.33% on four fruit and vegetable datasets, respectively, outperforming state-of-the-art methods (MobileViTV2, MixNet, MobileNetV2). The maximum memory usage during training is only 6.48GB, which is merely 13.8% of that required by state-of-the-art methods(MobileViTv2). The fruit and vegetable recognition model proposed in this study offers a more profound and effective solution, providing valuable insights for future research and practical applications in this domain. The code is available at https://github.com/Axboexx/CGViT.