Enhanced Neuromorphic Object Detection with Improved CNN and Flash Attention

Traditional object detection models exhibit limited efficacy in identifying spike camera images, frequently failing to accurately detect target objects with high confidence. They are also characterized by slow training and inference processes, coupled with suboptimal resource utilization. This study introduces an innovative method for the spike camera images object detection that leverages the "Flash Attention" mechanism, which has been effectively validated on the YOLO detection model. This method integrates a sophisticated self-attention layer into the model, significantly enhancing its capacity to process features of diverse sizes, shapes, and complexities. It effectively mitigates the deficiencies of current models in accurately recognizing the high-speed spike camera image under challenging lighting conditions, accelerating both training and inference times and optimizing resource utilization. Consequently, this advancement positions the model as a promising candidate for future applications in the high-speed spike camera image object detection.Our Flash Attention-augmented detection model achieved a notable mean Average Precision (mAP) score of 88% on the spike camera image dataset, marking a 31% improvement in training speed and a 21% increase in inference speed over the baseline model.