TomFusioNet: A tomato crop analysis framework for mobile applications using the multi-objective optimization based late fusion of deep models and background elimination

Abstract

Crop disease is a critical concern for the farmers which needs to be addressed to mitigate production loss. Multifarious frameworks have been proposed for the rapid recognition of tomato crop diseases, however, their practical deployment is a daunting task. The existing models suffer from performance issues such as overfitting and gradient vanishing. To alleviate these issues, this research proposes an end-to-end tomato crop analysis framework, TomFusioNet, for mobile applications. For feature extraction, the late-fusion technique is leveraged by aggregating the results of modified cross-domain transfer learning models using the maximum likelihood prediction strategy. Multi-layer Perceptron models are utilized as meta-learners. TomFusioNet’s pipeline comprises two modules, namely, DeepRec and DeepPred. DeepRec provides preliminary disease recognition results while DeepPred further identifies the type of illness in crop. Logic gate mapping is employed to reduce unnecessary wastage of mobile computation. The hyperparameters of DeepPred are tuned using the multi-objective optimization based non-dominated sorting genetic algorithm II for performance enhancement. We highlighted the significance of feature relevancy; therefore, a hue, saturation and value (HSV) color model-based background elimination technique is also proposed. TomFusioNet can be incorporated in a smartphone app, conceptualized for remote crop monitoring. The proposed DeepRec and DeepPred models achieved an average accuracy of 99.93% and 98.32%, respectively. According to experimental results, TomFusioNet, which is built on NSGA-II, beats state-of-the-art models with a convergence loss of 0.021 and an AUC value of 99.10%. The proposed app framework’s functioning does not require any human in loop. Furthermore, the latency of framework in real-time is less than 2 s, hence, it is effective for rapid tomato crop analysis.