Zusammenfassung
In recent years, deep convolutional neural networks (DCNN) have evolved significantly in order to demonstrate remarkable performance in various computer vision tasks. However, their excess storage requirements and heavy computational burden restrict their scope of application, particularly on embedded platforms. This problem has motivated the research community to investigate effective approaches ...
Zusammenfassung
In recent years, deep convolutional neural networks (DCNN) have evolved significantly in order to demonstrate remarkable performance in various computer vision tasks. However, their excess storage requirements and heavy computational burden restrict their scope of application, particularly on embedded platforms. This problem has motivated the research community to investigate effective approaches that can reduce computational burden without compromising its performance. Filter pruning is one of the popular ways to reduce the computational burden, where weak or unimportant convolutional filters are eliminated. In this paper, we propose a novel approach for filter pruning based on an adaptive multi-objective particle swarm optimization (AMPSO) to compress and accelerate DCNN. The proposed approach searches for an optimal solution while maintaining the trade-off between network's performance and computational cost. Extensive experiments on TernausNet and U-Net for high-resolution aerial image segmentation tasks demonstrate the superiority of AMPSO in finding a compact network model.
Altmetric