Lightweight image classification method and system based on neural network pruning

The invention discloses a lightweight image classification method and system based on neural network pruning, and the method employs the absolute value of a BN layer scaling factor as an index for measuring the importance degree of a convolution kernel, and introduces l1 regularization sparseness into the scaling factor in a loss function, thereby enabling a network to correctly recognize the importance degrees of different convolution kernels. By analyzing and calculating the convolution kernel change rate, the pruning proportion pbest% containing all convolution kernels with redundant parameters is determined, and therefore subsequent pruning operation is carried out. According to the invention, the image classification precision is improved. 本发明公开了一种基于神经网络剪枝的轻量级图像分类方法及系统，利用BN层缩放因子的绝对值作为衡量卷积核重要程度的指标，并在损失函数中对缩放因子引入l1正则化稀疏，使得网络能够正确识别不同卷积核的重要程度。通过分析和计算卷积核变化率，确定包含全部存在冗余参数的卷积核的剪枝比例pbest％，从而进行后续剪枝操作。本发明提高了图像分类的精度。