Fibonet: A Light-weight and Efficient Neural Network for Image Segmentation

In recent years, accurate models for image segmentation have become larger and more complex. However, it is hard to apply them into embedded devices which usually have limited space for data, energy and computing. Meanwhile, since many embedded devices do not support modifications of computing units, simple models with such requirements cannot be adapted to these devices. To achieve high accuracy for image segmentation in embedded devices, we propose a light-weight and efficient neural network, named Fibonet. Fibonet is constructed by cascading two Fiboblocks, with the Fibonacci structure which adjusts the combination of basic computing units through skip connections and feature reusing so that it is less computationally intensive and dataset-demanding. The experiments demonstrate that Fibonet can be embedded to the mobile terminal for real-time segmentation and can effectively balance accuracy and computing resources. Compared with Resnet18, Fibonet achieves 10.8% higher accuracy performance (mIoU: 0.533 vs. 0.481) with 84.4% fewer parameters (0.441M vs. 2.82M) using similar model width and depth.