Depthwise separable convolution architectures for plant disease classification

Abstract

Convolutional neural network has a huge partake and is still a dominating tool in the field of computer vision. In this study, we introduce a model with depthwise separable convolution architecture for plant disease detection based on images of leaves. We present two versions of depthwise separable convolution comprising two varieties of building blocks. Training and testing of the models were performed on a subset of publicly available PlantVillage dataset of 82,161 images containing 55 distinct classes of healthy and diseased plants. These depthwise separable convolutions achieved less accuracy and high gain in convergence speed. Several models were trained and tested, of which Reduced MobileNet achieved a classification accuracy of 98.34% with 29 times fewer parameters compared to VGG and 6 times lesser than that of MobileNet. However, MobileNet outperformed existing models with 36.03% accuracy when testing the model on a set of images taken under conditions different from those of the images used for training. Thin models were also introduced, which showed effective trade-off between latency and accuracy. The satisfactory accuracy and small size of this model makes it suitable for real-time crop diagnosis in resource constrained mobile devices.