First Layer Optimization of Convolutional Neural Networks for IoT Edge Devices

Optimizing Convolutional Neural Networks (CNNs) for hardware deployment has gained significant attention as the number of edge devices continues to grow. This is driven by the realization that edge computation reduces bandwidth usage and, consequently, overall power consumption. The first CNN layer, in particular, is receiving increased focus due to its direct impact on the network's overall accuracy. This paper provides an overview and summary of various first-layer optimization techniques. Additionally, a Hardware-Aware Technique (HAT) is proposed and evaluated for its potential as a first-layer optimization method. A comparison between the proposed HAT and other leading techniques, based on overall network accuracy and the quality of first-layer feature maps, demonstrates that the HAT is a strong candidate for first-layer CNN optimization. In a custom CNN architecture using a dataset acquired by a CMOS image sensor (CIS), the proposed HAT achieves a validation accuracy of 96.49%, which is highly competitive with other state-of-the-art approaches.