The Image Recognition and Classification Model Based on ConvNeXt for Intelligent Arms

Image recognition is typically a multi-feature decision problem. After performing certain preprocessing on images, various feature information is extracted to identify objects and scenes within the image. Image classification, which is usually based on image recognition, involves categorizing the image into predefined classes, and it is a highly challenging task. Both image recognition and classification face challenges due to the complexity of image features, limited computing power and efficiency. Traditional models have low accuracy, high computational complexity, and average generalization ability, while the new ViT models have high resource requirements and are prone to overfitting on small datasets. Starting with 700 images (including multiple angles of 15 different categories), this paper uses the ConvNeXt model to address the image recognition and classification problem for small intelligent robotic arms. This model is based on traditional CNN architecture, optimized by incorporating the advantages of Transformer. The method first preprocesses the images, then uses convolutional layers to embed the image into a high-dimensional space, encodes the features in this high-dimensional space, and then decodes the encoded features with convolutional layers. Finally, a cross-entropy loss function is used to optimize the classification task. Experiments show that the transfer learning ConvNeXt model, compared to other models, has lower computational complexity, lower resource requirements, stronger local perceptual abilities, and better small object feature recognition and long-term dependency handling capabilities. It can train effective features without requiring large amounts of data. This model has shown great results in assisting small intelligent robotic arms to recognize different objects and perform corresponding tasks, demonstrating the feasibility of the proposed method.