Lightweight recognition for multiple and indistinguishable diseases of apple tree leaf

Detection technology has been widely used for apple leaf diseases and pests in recent years, with the rapid development of deep learning. However, most of the existing identification equipment is deployed in the field, with limited network conditions and computing resources, whereas, a wide variety of diseases and pests, with limited recognition accuracy, which is not conducive to the development of digital agriculture. In this study, a fast identification (S-DenseNet-E) was proposed to realize the best effect on public data sets. The DenseNet's Dense module was selected to add an auxiliary model, in order to facilitate the identification of indistinguishable diseases. The S-DenseNet model was designed for the multi-diseases using the DenseNet121. The aggregation connection was adopted only in the output layer of the module in the S-DenseNet. The high reasoning speed of DenseNet121 was achieved to solve the dense connection. The average all-category F1-score of S-DenseNet was 85.14%, the inference time was 33.03 ms, and the number of parameters was 1.1 M. Compared with DenseNet121, the average F1-score of all categories was improved by 0.13 by 4.28, the inference time was shortened by 32.87 ms, and the number of parameters was reduced by 6.96 M. An auxiliary model was used to assist S-DenseNet to identify the indistinguishable disease, which was built S-DenseNet-E of composite recognition framework. Furthermore, the one-vote strategy was adopted under two-model voting. The S-DenseNet-E maintained the high recognition of the S-DenseNet for the multi-diseases, while the effective recognition of the indistinguishable disease indicated a low inference time and small parameters. S-DenseNet-E achieved an average F1-score of 85.86% in all categories, an F1-score of 70.10% for indistinguishable disease, an inference time of 38.92 ms, and a parameter of 2.2 M. Compared with the S-DenseNet, the S-DenseNet-E improved the F1-score by 4.28 on indistinguishable diseases, and the inference time by 5.89 ms. Compared with the DenseNet121, S-DenseNet-E shared an average F1-score improvement of 0.85 percentage points in all categories, an F1-score improvement of 5.18 percentage points in indistinguishable disease, a reduction in the number of parameters by 5.86 M, where the inference time was reduced by 26.98 ms. Therefore, the S-DenseNet-E presented better recognition performance for two complex situations, namely, apples suffering from multiple diseases and unidentifiable diseases, where was required fewer computational resources. The practical application of the model was also verified in the field. Specifically, 1730 apple leaf disease images were collected on site in Banan District, Chongqing, China. Three types were divided into healthy, Scab, and rust leaves. The improved model was compared with other models for experiments. The S-DensseNet model showed a recognition accuracy of 91.91% for healthy apple leaves, 91.83% for Scab apple leaves, and 98.43% for Trust apple leaves. The S-DensseNet model also demonstrated the least inference time, only requiring 42.72 ms. The experimental results show that the S-DenseNet and S-DenseNet-E can be expected to run on embedded edge computing devices, while more accurately identifying the multiple diseases and unidentifiable diseases of apple leaves, and fully meeting the actual production needs of apple orchards. The finding has an important significance for the development of digital agriculture.