Towards image-based measurement of accurate apple size and yield using stereo vision cameras

Accurate measurement of fruit size in apple orchards before harvest can have important implications on profits and management practices. Obtaining a distribution of fruit size can be labor intensive for a large orchard and therefore requires an automated system that can quickly and accurately size fruit on every tree. This study proposes an automated imaging system that uses stereo vision for finding the metric surface area of apples. Deep convolutional neural network models were utilized to classify apples as ideal candidates for sizing based on their orientation and visibility in an image. The results produced a correlation of apple size to apple weight of R²=0.69 making the system capable of capturing variability in fruit diameter distribution that ranges by 1cm (or equivalently 60 grams). There was also an improvement in correlation to yield when combining fruit size with fruit count than when utilizing fruit count alone.