用基于遗传算法的BP神经网络识别牛肉肌肉与脂肪

利用遗传算法的全局搜索能力,改进标准BP算法随机选取初始权重的不足,并构建了3-12-1的三层遗传BP神经网络,进行了3次牛肉肌肉与脂肪像素的分类试验,研究用BP网络对牛肉肌肉与脂肪两类像素点进行识别的可行性。以像素点的RGB值作为BP网络输入向量,每次训练集样本数62,测试集样本数43。测试的最终结果为：训练集的样本识别率分别为100%、100%、98.3871%;对应测试集的样本识别率分别为97.6744%、97.6744%、100%。试验结果表明,尽管基于遗传算法的BP神经网络对训练样本集以及测试样本集的肌肉和脂肪的识别率均在97%以上,但由于牛肉图像像素值在颜色空间中比较分散,有利于聚类的规律性不明显,因而是否可用BP网络来完成肌肉与脂肪的识别,还需要在网络拓扑结构、训练样本集等方面进一步研究。

Classifying beef muscle and fat pixels using BP-GA neural network

Genetic Algorithm(GA), which has the potential to search optimal global answer values of problem, was applied to improve the shortage of the original Back-Propagation(BP) algorithm with randomly selecting values in some limitations as its first weights. For the sake of discussing the feasibility of classifying beef muscle and fat pixels with BP neural network, a 3-12-1 three-layer BP-GA neural network was built, and three experiments on classifying the beef muscle and fat pixels were implemented with it. The input vector of the BP neural network is the RGB values of each pixel, and the training and test set samples are 62 and 43, respectively. The final results of the three experiments are as follows: the recognition rates of training samples are 100%, 100%, 98.3871%, respectively, and the recognition rates of test samples are 97.6744%, 97.6744%, 100%, respectively. The experimental results show that, although the recognition rates of the training sample sets and the test sample sets are all over 97%, due to the RGB values of beef image pixels being dispersive in color space and their characteristics available for clustering being inconspicuous, whether it can use BP neural network to classify beef muscle and fat pixels or not needs to study furthermore, such as constructing a rather reasonable network topology, or selecting rather appropriate training sample sets.