Tobacco Open-set Recognition Based on Feature Selection Reconstruction

The intercropping of mainstream cultivated tobacco varieties with unknown tobacco varieties significantly affects the quality and yield of tobacco. Conventional methods for identifying tobacco heavily depend on manual recognition. However, the resemblance in characteristics among different tobacco breeds restricts the differentiation between known and unknown varieties. To mitigate this challenge, the present study introduces the Cosine Feature Reconstruction Network (CFRN) for distinguishing between known and unknown tobacco breeds. The model employs a feature reconstruction selection module to extract fine-grained image features with disparities and integrates a dynamic information integration model to enhance the correlation between these features across various scales. Subsequently, the model further refines its decision boundary by integrating it with the cosine classifier. Furthermore, during the inference phase, we integrate it with a post-scoring function known as Decoupled Maximum Log Probability (DML). Experimental assessments conducted on Tobacco datasets and CUB dataset illustrate that CFRN surpasses prevailing open-set scoring methods in the majority of instances.