The Image Sharpness Estimation and the CNN Training Enhancement in the Empty Containers Recognition System of Reverse Vending Machine

Abstract

The automatic reverse vending machine (RVM) “Sortomat” accepts plastic bottles for further recycling. The analysis of the received containers is performed by running the neural network script. Computations are performed by the Raspberry Pi whose computing power is small and image processing by neural networks takes a lot of time. This paper discusses two procedures that verify the necessity to run a neural network script. The first function allows us to find out whether the camera is powered on and whether pictures are taken in focus and are sharp. The second function reports whether there is an object inside the RVM which is suitable for recognition. This approach helps to decrease the total operating time by estimating the necessity of neural network running and by avoiding the blurred and faulty image processing. The second problem discussed in this article is the image data source augmentation methods for object recognition accuracy enhancement.