Efficient Detection Model Using Feature Maximizer Convolution for Edge Computing

Deep learning is heavily influenced by the quantity and quality of data. Moreover, most deep learning models are developed and tested on servers equipped with high-performance GPUs and large memory capacities. However, for practical application in industrial fields or real-world scenarios, optimal performance must be achieved using limited resources and equipment. Therefore, this paper proposes the FMC (Feature Maximizer Convolution). This method aims to enhance performance with limited data by extracting as diverse features as possible and assigning more weight to crucial feature maps, which are then passed on to the next layer. Additionally, to ensure real-time performance on limited hardware, DSC (Depthwise Separable Convolution) is employed instead of standard convolution to reduce computational load. The approach is applied to deep learning models on datasets such as COCO, VisDrone, VOC, and xView, and its performance is compared with existing networks. Inference experiments are also conducted on the edge device Odroid H3+. The proposed network shows a 30% average reduction in the number of parameters compared to existing networks and a 5% increase in inference speed. On the Odroid H3+, the inference speed improved by an average of 2.5 ms, resulting in an increase from 19 FPS to 20 FPS.