Abstract: Fish with different weight have different feeding ability. In aquaculture, accurate estimation of fish weight is beneficial for feeding fish in proper amount, avoiding waste of feed and water pollution. Firstly, the image and weight data of crucian carp were collected with a coin as a reference. Secondly, the image was preprocessed and the features of crucian carp and coin were extracted. Finally, BP neural network, Elman neural network and neural network constructed by Numpy library were used to realize the multi-feature weight estimation of crucian carp. The results showed that the determination coefficients of BP and Elman neural network were 0.925 6 and 0.906 4, the mean square errors were 0.003 68 and 0.004 55. The determination coefficient of neural network estimation constructed by Numpy library was 0.823 7, and the mean square error was 0.008 1. Therefore, the BPNN-area-perimeter and Elman-area-perimeter methods could quickly and accurately estimate the weight of fish, realize the precise feeding of fish in aquaculture, classify the fish according to the weight after the fish was caught, and promote the progress of fishery modernization.

Key words: weight estimation, neural network, multi-feature, aquaculture, image process

摘要： 质量不同的鱼摄食能力不同，准确估计鱼体质量有利于水产养殖中鱼类精准投喂，避免饲料浪费及水体污染。首先使用1元硬币作为参照物采集鲫鱼图像和体重数据，其次对图像进行预处理，提取鲫鱼和硬币的特征值，最后采用BP神经网络、Elman神经网络以及Numpy库构建的神经网络实现多特征的鲫鱼质量估计。结果表明：使用BP和Elman神经网络估计鲫鱼质量时决定系数分别为0.925 6和0.906 4，均方误差分别为0.003 68和0.004 55。采用Numpy库构建的神经网络估计时决定系数值为0.823 7，均方误差值为0.008 1。因此，使用BPNN-面积-周长和Elman-面积-周长方法能够快速、准确地估计鱼体质量，实现水产养殖中鱼类的精准投喂，以及在鱼类被捕捞后根据质量进行分级，推进渔业现代化的进展。

关键词: 质量估计, 神经网络, 多特征, 水产养殖, 图像处理