基于PSO-ConvK卷积神经网络的肺部肿瘤图像识别
Lung tumor images recognition based on PSO-ConvK convolutional neural network

针对卷积核随机初始化以及梯度下降法训练卷积神经网络易陷入局部最值问题,提出粒子群算法优化卷积核(particle swarm optimization-convolution kernel, PSO-ConvK)的图像识别方法。使用参数迁移法构造卷积神经网络,并提取卷积核,利用PSO不断更新粒子的速度和位置,寻找全局最优值以初始化卷积核,将其传递到卷积神经网络,用肺部肿瘤数据训练卷积神经网络,结合梯度下降法修正网络权重,使得PSO算法的全局优化能力与梯度下降法的局部搜索能力相结合。试验通过批次大小、迭代次数以及网络层数3个角度验证方法的有效性,并与高斯函数优化卷积核进行对比。结果显示, PSO优化卷积核的识别率始终高于随机化卷积核和高斯卷积核,识别率最终达到98.3%,具有一定的可行性和优越性。
In order to solve problems that convolution kernel was random initialization and the gradient descent method to train convolution neural network was easy to fall into local minimum, an image recognition method based on particle swarm optimization for convolution kernel was proposed. CNN(convolution neural network) was constructed by using the parameter migration method, and convolution kernel was extracted. The particle swarm algorithm was used to update the velocity and position of particles constantly and find the global optimal value to initialize convolution kernels. Convolution kernels were transferred to convolution neural network, and lung tumor images were used to train them. CNN model was trained by lung tumor images, and gradient descent method was used to modify network weights, hence global optimization ability of PSO algorithm was combined with local search ability of gradient descent method. The experiments verified effectiveness of method through three perspectives: batch sizes, iteration numbers, and network layer numbers. The particle swarm algorithm was compared with gauss function. The recognition rates of PSO optimized convolution kernel were always higher than that of randomized convolution kernel and gauss convolution kernel, its recognition rate reached 98.3%, which had certain feasibility and superiority.