Lite-SRGAN and Lite-UNet: Toward Fast and Accurate Image Super-Resolution, Segmentation, and Localization for Plant Leaf Diseases

Complex deep convolutional networks are typically designed to achieve state-of-the-art results. Such networks require powerful computing resources and cannot work efficiently on resource-constrained devices particularly for real-time use. To address these challenges, this study introduces resource-efficient lightweight approaches for segmentation, localization, super-resolution, and classification tasks. On this basis, we propose two novel lightweight architectures named: Lite-UNet and Lite-SRGAN. We validated the effectiveness of our proposed networks using the large publicly available Plant Village dataset. Lite-UNet network is used for performing segmentation and localization tasks, while Lite-SRGAN network is used for performing the super-resolution task. The proposed Lite-UNet outperforms U-Net with slight gains of 0.06% and 0.12% for dice coefficient and Intersection over Union (IoU) respectively while achieving significant reductions of 15.9x, 25x, and 6.6x in terms of parameters, floating-point operations per second (FLOPs), and inference time respectively. In addition, the proposed Lite-SRGAN achieves comparable qualitative and quantitative results compared to SRGAN with significant reductions of 7.5x, 7.8x, and 2.7x in terms of parameters, FLOPs, and inference time respectively when upsampling the low-resolution images from $64\times 64$ to $256\times 256$ (4x upscaling). Similarly, it achieves a reduction of 7.1x, 11.2x, and 1.9x when upsampling from $128\times 128$ to $256\times 256$ (2x upscaling). For classification purposes, a two-stage classification approach is introduced, in which the crop species and their leaf diseases are recognized respectively. Different models are utilized in both stages including MobileNetV3, DenseNet121, and ConvNeXt. The best accuracy obtained on the testing set is 99.76% when using the proposed methods together, which outperforms several other related studies. Source code is available at https://github.com/hosamsherif/LiteSRGAN-and-LiteUNet