Size and mass prediction of almond kernels using machine learning image processing

Abstract

After harvesting almond crop, accurate measurement of almond kernel sizes is a significant specification to plan, develop and enhance almond processing operations. The size and mass of the individual almond kernels are vital parameters usually associated with almond quality, particularly head almond yield. In this study, we propose a novel methodology that combines image processing and machine-learning ensemble that accurately measures the size and mass of whole raw almond kernels (classification -Nonpareil) simultaneously. We have developed an image-processing algorithm using recursive method to identify the individual almond kernels from an image and estimate the size of the kernels based on the occupied pixels by a kernel. The number of pixels representing an almond kernel was used as its digital fingerprint to predict its size and mass. Various popular machine learning (ML) models were implemented to build a stacked ensemble model (SEM), predicting the mass of the individual almond kernels based on the features derived from the pixels of the individual kernels in the image. The prediction accuracy and robustness of image processing and SEM were analyzed using uncertainty quantification. The mean error in estimating the average length of 1000 almond kernel was 3.12%. Similarly, mean errors associated with predicting the 1000 kernel mass were 0.63%. The developed algorithm in almond imaging in this study can be used to facilitate a rapid almond yield and quality appraisals.

Face Recognition Using Sf3CNN with Higher Feature Discrimination

“…We have compared our results with the work on face recognition by Mishra et al [10]. In the work by Mishra et al [10], the experiment has been performed on CVBL database using 3D residual networks when the loss function is Crossentropy.…”

“…We have compared our results with the work on face recognition by Mishra et al [10]. In the work by Mishra et al [10], the experiment has been performed on CVBL database using 3D residual networks when the loss function is Crossentropy. Both, in our experiment and in the experiment performed by Mishra et al [10], a clip of temporal length 16 is input to 3D residual network of different depths and genres.…”

“…We performed the experiments on CVBL (Computer Vision and Biometric Lab) database [10] using Sf 3 CN N framework. The optimizer used is Adamax.…”

“…The compact feature generated from 3D CNN contains more information than that generated using 2D CNN. This allows for an efficient and robust face recognition [10]. The features generated using 3D CNNs however, are not highly discriminative and hence affect the accuracy [10].…”

“…This allows for an efficient and robust face recognition [10]. The features generated using 3D CNNs however, are not highly discriminative and hence affect the accuracy [10]. The discriminative ability of the loss functions has been increased by using the concept of angular margin.…”