Counting apples from the Minneapple dataset

Made in colaboration with Emil Delvaux

My first proper crash into neural networks.

Segmentation using color channels

Simple segmentation based on brightness components.

Images are split, equalized, thresholded and filtered with morphological operations before separated into clusters.

• Using RGB color channels, red channel used as base, and others extracted.
• Using HSV color channels, using brightness component.
• Using Both composite approach, using best of the previous.

Basic CNN implementation from scratch in two parts:

• Design and training of the CNN, using the counting dataset from Minneapple. Structure and weights saved for later.
• Using the CNN. Splitting the image into chunks, feeding each one into the trained network and summing up the total of apples.

Verdict

Overall results are... not the greatest, quite bad actually with an error of around +-30% from truth.

As a learning experience however, the project was extremely rewarding, allowing us to crash headfirst into traditional and NN segmentation techniques.