Hardware Oriented Strip-wise Optimization (HOSO) Framework for Efficient Deep Neural Network

Machine learning methods combined with other computer science technologies have exhibited great potential and unique opportunities to enhance computer vision tasks, offering an effective alternative and a viable method to image processing studies. However, current state-of-the-art object detection methods are either accuracy-oriented within a large model or speed-oriented within a lightweight model. Although some techniques are proposed to compress and accelerate neural networks, most algorithms mainly aim at achieving optimal goals on accuracy and model size based on evaluation in software and lead to a lack of consideration of hardware platforms.To tackle this challenge, in this work, we propose an Hard-ware Oriented Strip-wise Optimization (HOSO) framework, an efficient method that can consider target hardware platforms to compress the machine learning model while retaining high accuracy. Compared with the existing pruning methods, the proposed HOSO framework focuses on Hardware-oriented optimization based on strips instead of the filter or weight. In addition, the proposed HOSO framework converts the standard convolution into a strip-wise convolution. Thus, a novel hardware implementation is proposed. Experimental results show that our proposed HOSO framework and its hardware implementation can compress the model into a smaller size while retaining high accuracy.