Classification for Tomato Disease with Imbalanced Samples Based on TD-MobileNetV2

Image classification based on deep learning requires a large number of training samples. The main obstacle, from a practical perspective, has been difficult sample obtaining of several disease categories. The key task of this paper is to solve the problem of bad accuracy performance and low robustness influenced by imbalanced dataset. In the stage of image pre-processing, weighted random sampling algorithm and data augmentation techniques are applied to match up different category samples consistent. A CNN architecture based on MobileNetV2 and targeted dropout algorithm is proposed to improve the robustness of the conventional model. Results demonstrate that the proposed model achieves a performance very close to that of conventional CNN architecture. Moreover, the detrimental effects of Gaussian and Salt&Pepper noise is weakened, and the technique offers significant performance gain on complex background samples over the traditional deep learning methods.