Objective  An improved diagnostic model on apple leaf diseases was developed applying the lightweight residual networks.

  Method  Data imbalance and overfitting of the original ResNet18 model was reduced by offline and online enhancements on its generalization ability. A scaling factor was introduced to minimize the number of network parameters and maximize the pooling layer in the constant mapping of down-sampled residual structure instead of 1×1 convolution. Redundant features in pictures were eliminated, and sensory field of the model increased. The first 7×7 convolutional layer of ResNet18 was replaced with a multi-scale feature extraction module to enrich fine lesions extraction. Finally, the DenseBlock module was inserted in the network to fully utilize valid shallow features.

  Results  The improved ResNet18 model achieved an accuracy of 97.94%, which was a 0.88 percentage increase, with a significant 90.77% reduction on the program size of 3.97 MB. It performed superbly in comparison to other light-weight models, such as ShuffleNetv2, MobileNetv3, and EfficientNet, or the classical models, such as Inceptionv2, DenseNet, and ResNet.

  Conclusion   The improved ResNet18 model could accurately identify the apple leaf diseases under complex circumstances. With fewer parameters than the original program, it could be more easily installed in a variety of devices for the diagnosis.