Chilli Identification and Grading in pre/post-harvest Environment based on Computer vision and Deep Learning approaches

Abstract

Chilli, one of the spice produce, needs grading before being marketed for produce quality assurance. Manual chilli grading involves high labour cost, time-consuming, inconsistent, and expensive warranting technology intervention. In this work, a non-destructive approach to identify dry chilli images into three levels as good quality, medium quality and poor quality, using a deep learning architectures and grade them are adopted to reduce computation overload. The database of chilli grown in North Karnataka region is prepared as no standard chilli datasets are available. Dry chilli images dataset are augmented to train the dataset for transfer learning (DL) models, namely VGG16, ResNet and EfficientNet-D0 to analyse suitability of good model for the grading of chilli images. Further, work needs integration of the algorithm into automatic chilli grading tool. The proposed EfficentDet model is found suitable and yielded accuracy rate of 95.62% were in VGG16 and Resnet models accuracy was 82.67% and 83.88%. EfficientDet model out performs in terms of grading the dry chilli images.