43 / 2025-03-31 20:36:08

A Novel Model for Diseased Chicken Recognition Integrating Visual and Textual Features

Knowledge Graph,Muleimodal,ConvNeXt,Deep learning

3. 智慧牧场与动物福利 > （2）智慧养殖的大数据、机器学习和物联网应用

To address the challenge posed by the limited information derived from a single visual feature, which results in inadequate recognition of diseased chickens, the ConvNeXt model has been optimized through the integration of the MASA Attention Mechanism Module and the DeConv Convolution Module. This optimization enhances the model's capability to focus on key visual lesion features, such as color and texture, and improves feature reconstruction by employing multi-scale spatial adaptation and recovering spatial resolution through the fusion of multilevel feature methods, thereby increasing the model's recognition accuracy. Additionally, a comprehensive and fine-grained chicken disease knowledge graph has been developed. By employing a graph attention network (GAT), the nonlinear relationships between symptoms and diseases are captured, thereby augmenting the bimodal fusion model's understanding of semantic information. Furthermore, a cross-modal gated cross-attention mechanism (CGC) has been designed to dynamically adjust the weights of image and knowledge graph information. This mechanism enhances the overall recognition performance of the model by dynamically filtering cross-modal information and adaptively fusing multi-source information. The experimental results indicate that the recall, precision, and F1 score of the optimized model have been enhanced by 3.78%, 3.88%, 3.86%, and 3.89%, resulting in a recognition accuracy exceeding 95%.