FusionNet: a New Perspective of CNN for Image Classification

The civil livelihood plan is related to corn, or maize, as one of the most significant grain products that suffers from various kinds of diseases, resulting in a decline in production. Traditional manual examination for disease in terms of prevention and control is time-consuming with low efficiency. Therefore, this paper proposes a new architecture called FusionNet fully based on deep convolution network to improve the speed and accuracy of the classification of various maize diseases. Using focus splitting, identity shortcut, and multi-feature channel fusion, this model presents a new consideration of larger kernels utilized in concatenated for features that offer more effective receptive fields to facilitate the reduction of reused redundancy, especially for middle-layer abstract information, which could compensate for the shortage of feature selectivity of CNN. Moreover, the total number of parameters is limited by depth-wise(DW) convolutions. The validity of the model is demonstrated on modified PlantVillage and PlantDoc datasets, achieving an accuracy of 94.9% on average.