Vision-knowledge-fusion-based agricultural pest recognition and intelligent auto-labeling

The present study proposes a novel methodology for the classification and recognition of agricultural pests. This methodology is based on the development of a complex network architecture that combines an enhanced Transformer model with an adaptive convolutional neural network (CNN). The enhanced Transformer incorporates a dynamic attention mechanism and multi-scale feature extraction strategy, enabling it to adaptively adjust the attention region based on image content and effectively capture multi-scale information. This addresses the challenges of image diversity and noise interference in agricultural pest recognition. The proposed network aims to leverage the strengths of both models to enhance feature extraction and improve the accuracy and robustness of pest classification. The enhanced Transformer model incorporates a dynamic attention mechanism and multi-scale feature extraction strategy, enabling it to adaptively adjust the attention region based on image content and effectively capture multi-scale information. This addresses the challenges of image diversity and noise interference in agricultural pest recognition. The experimental results obtained from the implementation of the proposed methodology have been shown to validate its effectiveness, with the multi-layer CNN demonstrating superior performance in terms of accuracy, recall, and F1 score when compared to other models. The integration of the enhanced Transformer and adaptive CNN has been demonstrated to significantly enhance the model's capacity to extract intricate image features while maintaining robustness under diverse environmental conditions. The ensemble model with weighted voting has been demonstrated to enhance the accuracy of classification, thereby underscoring the efficacy of this methodology in the recognition of agricultural pests. In summary, this study provides a novel and effective technical solution for intelligent control of agricultural pests, demonstrating high application potential in real agricultural environments. A multimodal dataset was constructed, combining pest images and knowledge data. In addition, a novel architecture was proposed, integrating an enhanced Transformer with dynamic attention and multi-scale extraction, alongside an adaptive CNN that adjusts convolution kernels dynamically. This combination has been demonstrated to enhance the extraction of features and the adaptability of models in a variety of agricultural contexts. A multimodal intelligent labeling method was presented, which synergises visual and knowledge data, thereby enhancing classification accuracy through weighted voting. The method has been demonstrated to reduce manual annotation errors and improve decision-making by fusing expert knowledge with model predictions. Through extensive experimentation and the execution of ablation studies, the model attained an accuracy of 95.3%, surpassing the performance of existing techniques in terms of accuracy, recall, and robustness. This demonstrates the methodology's potential for practical agricultural pest recognition and intelligent control applications.