Tiny defects identification of mechanical components in die-cast aluminum using artificial neural networks for micro-controllers

This work was targeting the identification of tiny visual defects which were present on mechanical components in die-cast aluminum, just produced by a foundry. From those, an image data-set has been created composed by 500 colored pictures, 2048x850 pixel resolution each. The tools to perform the analysis were based on Artificial Neural Network (ANN) to achieve accurate defects detection. Moreover we were interested to map resulting ANNs topologies on micro-controllers MCUs as well as investigating architectures for efficient mapping. At first we have designed an ANN without tight implementation constraints, memory and computational, achieving 86% accuracy. Careful analysis of memory and computational needs, by using X-CUBE-AI tool, concluded it required an embedded RAM memory of 786 KB combining 64x64 input and activation buffers. To lower that number we explored an ANN implementation by removing the residual connections between convolutional layers featuring stride 2. That version achieved an accuracy of 75% in 100KB of embedded RAM.