Computer Vision – Fruits Recognition
Aim of the Project
● Create an AI model to detect and recognize strawberries on an image from a camera
● Develop a web service & API for the AI model integration and usage
● Accuracy 84%
● Image processing time <0.5 sec
● AI Architecture: Faster RCNN (backbone resnet 50)
● For ML: TensorFlow, Keras, Python.
● For Backend: Flask, Python.
● For Environment: Docker.
The development was made from scratch using 2500 marked up images with defined 6 states of berries. During R&D phase we have compared Deeplabv3+ Gaussian and Faster RCNN (backbone resnet 50), the last approach gave a bit better results.
The model is able to work with the following defects on images:
– partly blurred images
– bright glares from the sun
The model can detect 6 states of a berry:
– small green berry
– green berry
– white berry
– start colouring berry
– ripe berry
Initial model accuracy was increased by increasing the number of learning dataset and quality of markup.