Fruit Detection and Segmentation Using Customized Deep Learning Techniques

Abstract

Farmers may make better crop management decisions with accurate yield estimation. The key challenge for accurate fruit yield estimation is to distinguish and pinpoint the fruit from the tree in the field. In this model, we have used the U-Net architecture to cope and investigate this challenge. U-Net is based on sementic segmentation used for object detection and localisation. U-Net's contraction path encodes and extract the features from the source object (mango pictures), while the expansion path decodes the image by recovering the resolution for better localisation. This study focus on mango fruit an we employed the ACFR Mango Dataset. The all dataset images was divided into three classes: train, validation, and test images. The constructed model evaluated with test iamges that were not part of the training. Our model predicted accuracy and test image loss both were 98.66% and 0.0268%, respectively.