Computer Vision News - January 2019

Also, the packaging house needs to have control over the final grade of the produce as well as each sub-grade, and these controls need to be modified often without re-training or even stopping the system, which would perturbate the operations at the site. One example of issue that we detected and solved : differentiating the stem and blossom areas of the fruit (which do not downgrade the fruit) from scars and blemishes on the fruit surface (which do). In that way, when grading the spots in that area, stem and blossoms are ignored. Using one step that will give one grading answer to each input image, might be very problematic: it would need a lot of retraining when the conditions of the field change and phenomena like plant pests and diseases appear in the data. Thus, RSIP Vision and Sunkist RTS developed this stepwise approach based on AI that finds the specific features that define the quality of the fruit and provides a report for each specific area . The final score is obtained by aggregating each individual score, thus enabling to grade the produce according to market requirements. The Deep Learning method has to be adapted according to every specific feature and phenomena, so that we are using the optimal network: the one that gives the best accuracy result in the detection task of the relevant features - stains, diseases, stems and blossoms. Once adapted, we train the system accordingly to obtain the most robust results , that will not change during the season, even when ripeness of produce is changing. This approach gave Sunkist Growers exceptional results, enabling the successful detection of new phenomena, the control of the result and the analysis of any issue arising during operation. The great novelty of this system introduced by Sunkist RTS and RSIP Vision is the ability to utilize state-of-the-art AI in a robust and controlled setting . The outcome is very improved results with respect to the traditional method. Project 13 Computer Vision News A project by RSIP Vision Take us along for your next Deep Learning project! … very improved results with respect to the traditional method.