The purpose of this paper is to offer a machine vision approach for classifying cocoa beans based on their morphological properties. Using traditional machine learning approaches, the shape and size of cocoa beans were retrieved from photographs. A series of image processing techniques are used to extract the features from the photos. Finally, typical machine learning approaches such as KNN, SVM, Decision Tree, and Random Forest are used to divide the cocoa beans into four groups: large, medium, small, and rejected. A comparison of different methodologies is also carried out. Two optimization strategies, Univariate Selection and Feature Importance, are used to maximize retrieved features prior to training the model. For performance analysis, trained models are evaluated using stratified K-fold cross validations and the mean cross validation score is produced. The Random Forest Classifier has the greatest accuracy score of 0.75, according to the results of the experiments. Keywords: Cocoa beans, Classification, Image processing, Machine Learning, Feature Optimization.

Physical and biochemical attributes are commonly used for characterization of rice. The physical attributes are related to the quantification of size, shape, colour and texture of the rice grains. Biochemical attributes are assessed from cooking and eating characteristics of rice and are termed like alkali spreading value (ASV), amylose content (AC), gel consistency (GC), grain elongation etc. Estimation of biochemical attributes are often time consuming and require meticulous effort for sample preparation, storage and manual measurement. The gelatinization temperature (GT) is related to Alkali spreading value of rice and is partly associated with the amylose content of the starch. GT has a negative correlation with cooking temperature of rice. In this paper image analysis technique has been proposed for discrimination of rice. A portable flat bed scanner has been used as the imaging device and image analysis software has been developed to measure the rate of dispersion during ASV testing. This machine vision technique is a faster and effective way to determine the ASV. The results obtained are promising towards this new approach for objective estimation of ASV.

Turmeric ( Curcumina Longa ) is a globally traded commodity which is subjected to economically motivated chemically unsafe adulteration, namely metanil yellow. In this work, we report a simplistic and convenient approach to find the adulteration of turmeric with metanil yellow by near-infrared (NIR) spectroscopy coupled with chemometrics. Pure turmeric sample was prepared in the laboratory and spiked with different concentrations of metanil yellow. The reflectance spectra of 248 pure turmeric, metanil yellow, and adulterated samples (1–25%) ( w / w ) were collected using NIR spectroscopy. The calibration models based on NIR spectra of 144 samples were built for two different regression models, principal component analysis (PCR), and partial least square (PLSR) methods. Another 72 samples were used for external validation. The coefficient of determination ( R 2 ) and root mean square error of calibration for validation and prediction were found to be 0.96–0.99, 0.44–0.91, respectively, for most of the results depending upon different pre-processing techniques and mathematical models used. The original reflectance spectra, the 1st derivative plot, the plot of PLSR regression coefficient ( β ), and the first three principal component loadings revealed metanil-related absorption regions. To verify the robustness of the models, the figures of merit (FOM) of the models were calculated with the help of net analyte signal (NAS) theory. Overall, it was found that PLSR yielded superior results as compared to the PCR technique. These methods can be applied to other spices also to detect the adulteration rapidly and without any prior sample preparations and with low cost.