一种混合阈值剪枝的稀疏化训练图像识别算法

卷积神经网络在图像识别的应用中,有大量的冗余参数,增大了计算量,降低了网络运行速度.针对这个问题,提出了一种混合多阈值的稀疏化训练剪枝算法,通过稀疏化训练和混合全局与局部阈值的剪枝算法,压缩了神经网络的模型.通过对缩放因子L1正则化,使重要性低的通道值接近0,进行稀疏化训练,再引入全局阈值和局部阈值剪枝掉接近于零的通道并防止模型向粗粒度方向压缩,并对其进行训练微调参数,得到了混合阈值剪枝的精简模型.最后为了验证本文方法有效性,在DOTA(a large-scale dataset for object detection in aerial images)数据集中进行测试,该算法在小幅度降低图像识别精度的前提下,成功地压缩模型90％大小,加快了53％的计算速度,取得了较好的效果.

A Image Recognition Algorithm of Mixed Threshold Pruning and Sparsity Training

There are a lot of nuisance parameter in the application of image recognition of convolutional neural network, which increases the amount of calculation and reduces the speed of network operation. In order to solve this problem, a mixed multi threshold pruning algorithm has been put forward by this paper, which is to compress the model of neural network through sparse training and the printing algorithm with mixed global and local thresholds. Through the regularization of the scaling factor L1, the channel value of low importance was close to 0, sparse training has been carried out and then, the global threshold and local threshold were introduced to cut off the channel close to zero and prevent the model from compressing to the coarse-grained direction. Lastly, in order to verify the effectiveness of this method. The test has been carried out in DOTA remote sensing image data, Under the premise of reducing the accuracy of image recognition, the algorithm successfully compresses the size of the model and speeds up the calculation.