Precise identification of tomato leaf diseases based on the real background of DPEN

As an important economic crop, tomato is vulnerable to various diseases, and these diseases often have high visual similarity, making identification difficult. A delay in diagnosis can have a significant effect on tomato yields. Traditional manual visual inspection methods have poor accuracy, while laboratory diagnostic methods are inefficient, making them unsuitable for large-scale agricultural scenarios. To address this challenge, this study involved the collection and construction of a tomato leaf disease dataset in a real planting environment, and AutoAugment was used to achieve sample diversity and balance the number of training samples in different categories. Additionally, a dual-path ensemble network (DPEN) was proposed, which combines the multiscale feature extraction advantages of GoogLeNet with the dense connection mechanism of DenseNet121. The experimental results show that, compared with the comparison models, the DPEN achieves an identification precision of 98.80% on the self-built dataset, which is an improvement of 2.33% to 9.24%, and a reduction in the number of parameters by 7.09 M compared with GoogLeNet and 2.07 M compared with DenseNet121. The experimental results on public datasets further demonstrated the accuracy of the proposed DPEN model in identifying tomato leaf diseases in complex backgrounds. These results prove that the DPEN model can achieve precise, rapid, and efficient identification of tomato leaf diseases in complex backgrounds, providing technical support for smart agriculture applications.