Saffron is one of the costlier spices that are cultivated in specific regions of the world. Due to its restricted accessibility and more popularity, eventually saffron adulteration is one of the concerning issues in the recent times. It becomes difficult for human vision to discriminate between real and adulterated saffron samples. With the emergence of visual computing and data-driven algorithms, the saffron adulteration prediction systems (SAPS) are designed to predict the original and adulterated saffron samples. However, the majority of the techniques exhibit promising performance but the problem of generalization capabilities (unseen – samples) and scarcity of the saffron databases are still open research challenges. In this work, to overcome these issues, we propose a novel ensemble-based saffron prediction model (SaffNet) using statistical image features for the detection of contamination in the Kashmiri saffron. As data-driven approaches mainly rely on the representative samples, but to the best of our knowledge the standard benchmark datasets for Kashmiri saffron is not available. Therefore, we have created our novel Saffron dataset (Saff-Kash) collected afresh from different parts of Kashmir valley that includes the samples of both the authentic and adulterated saffron classes. The primary aim of the work is to anticipate the adulteration in saffron samples. Thereafter, these images are pre-processed and the dataset is prepared for the proposed SaffNet model. The SaffNet architecture designed using gradient boosting ensemble evaluated on Saff-Kash outperforms the outcomes of individual classifiers i.e., Support vector machine (SVM), decision tree, and K-Nearest neighbor (KNN) with an overall accuracy of 98%. Moreover, the execution time taken by the SaffNet model for training the SVM classifier is 8.56 milliseconds whereas for gradient boosting classifier it is 7.7 milliseconds.

No abstract