Abstract: In order to improve the effect of convolutional neural network on tomato disease image recognition, an improved multiverse optimizer algorithm was proposed. Firstly, the ports of wormholes were designed, and the ratio range between the number of ports of the white hole and the black hole was set to enable movement of the verse in multiple directions through wormholes. Secondly, an information transfer model of the verse was established based on the twoway motion, allowing for faster evolution and enabling the verse to obtain information from the optimal verse. Additionally, the expansion rate of the verse was corrected using nonlinear adjustment. Thirdly, the texture, color, and shape features of tomato diseases were extracted, and the activation and loss functions of the convolutional neural network were designed using the Spearman correlation coefficient to determine the optimal number of parameters in the multiverse. Finally, the algorithm flow was given. The experimental simulation showed that the proposed algorithm had a higher recognition accuracy compared to other algorithms, with an average value of 97.34% for bacterial spot, 97.03% for early blight, 9708% for late blight, 97.14% for leaf mold, 97.12% for leaf spot, and 97.20% for spider mite, while also having a lower consumption time than other algorithms.

Key words: multiverse, convolutional neural network, tomato disease, image recognition

摘要： 为提高卷积神经网络对番茄病害图像识别的效果，提出改进多元宇宙算法。首先设计、增加虫洞的端口，设置白洞侧的端口数量与黑洞侧的端口数量比值范围，使得宇宙的运动通过虫洞能够多方向进行；接着基于双向运动建立宇宙信息转移模型，宇宙能够正向、逆向进行的信息交流，加快宇宙的进化，使得宇宙都能够获得最优宇宙的信息，非线性调节对宇宙的膨胀率进行修正；然后对番茄病害提取纹理特征、颜色特征和形状特征，对卷积神经网络的激活函数与损失函数设计斯皮尔曼相关系数确定多元宇宙的优化卷积神经网络参数；最后给出算法流程。试验仿真表明：该改进算法对番茄各种病害识别正确率高于其他算法，细菌斑平均值为97.34%，早疫病平均值为97.03%，晚疫病平均值为97.08%，叶霉菌平均值为97.14%，叶斑病平均值为97.12%，蜘蛛螨平均值为97.20%，同时消耗时间小于其他算法。

关键词: 多元宇宙, 卷积神经网络, 番茄病害, 图像识别