Empirical Analysis of Squeeze and Excitation-Based Densely Connected CNN for Chili Leaf Disease Identification

Chili is one of the world's most extensively used spices. The changes in weather and environmental circumstances cause several diseases in chili leaves, which have an impact on yield of the spice. An optimized squeeze and excitation densely connected convolutional neural network (SEDCNN) architecture is proposed for early detection of chili leaf diseases. In the proposed research work, integration of SE block including pre-SE, post-SE, identity-SE, and standard-SE is individually performed and among these standard SE integrated with SEDCNN provides better performance. This architecture is evaluated with artificial neural network (ANN) conventional CNN, dense CNN, ResNet, ResNet with SE, separable standard SE, grouped standard SE, dilated standard SE, and nine different transfer learning networks. The standard-SE integrated with SEDCNN architecture has achieved higher disease identification accuracy of 97%. The empirical analysis of various block SEDCNN architectures and hyperparameter fine tuning resulted in the best accuracy with eight block SEDCNN, Adam optimizer with learning rate 0.001, batch size of 16, reduction ratio of 64, global average pooling squeeze operator, sigmoid excitation operator, Xavier normal kernel initializer, and categorical cross-entropy loss function. The proposed SEDCNN architecture integrated with standard SE identifies chili leaf disease with high accuracy, due to the incorporation of skip connections, SE blocks, and reduction ratio in the proposed architecture. The skip connection is utilized for feature forwarding, and the convolution layer weight update is being used for squeeze and excitation. Further, chili leaf disease identification accuracy is improved to 98.86% with augmentation.