摘要： 鱼苗数量的精准统计对提升苗种品质评价、养殖密度估算、鱼苗销售等环节的智能化水平具有重要作用。针对大菱鲆鱼苗个体小、透明度高以及体型不规则等影响计数精准度的问题，提出了一种基于ResNet34的大菱鲆鱼苗识别计数方法。首先，设计了一套适用于微小目标计数需要的图像采样装置，采用图像预处理方法实现鱼苗前景分割和初步定位。为了有效统一样本空间和待识别目标空间，利用最小外接矩规则化初步定位前景图像，构建图像样本集。大菱鲆鱼苗识别阶段，利用相同预处理方法获取待识别目标区域，并引入ResNet34模型作为识别模型实现待识别目标区域苗种识别；最后，通过统计所有待识别目标识别数量结果实现大菱鲆苗种计数。结果显示：本方法在微小鱼苗识别计数方面取得了较好的精度，利用ResNet34模型的大菱鲆鱼苗的识别平均准确率达到94.27%，比基于SVM方法（识别精度85.8%）和AlexNet（识别精度87.04%）方法识别精度分别提高8.64个百分点和7.4个百分点，优于ResNet18（识别精度93.21%）和ResNet50（识别精度93.83%）等相似结构的识别效果。本模型鱼苗计数的平均准确率达到96.28%。研究表明，提出的样本集构建和识别方法能够满足微小目标计数需求，可为鱼类苗种计数提供了技术借鉴。

关键词: 鱼苗计数, 图像识别, 大菱鲆, ResNet34模型

Abstract: The accurate statistics of the number of fry plays an important role in improving the intelligent level of fry quality evaluation, breeding density estimation and fry sales. At present, the counting of Turbot seedling mainly relies on manual labor, which is of high labor intensity and low precision. Because of the characteristics of turbot fry, such as small individual, high transparency and irregular shape, the traditional method of seed counting cannot be directly applied to Turbot seed counting. To solve the above problems, a method of identifying and counting turbot fry based on ResNet34 was proposed. Firstly, a set of image sampling device was designed, which consisted of camera, light source and container, etc., and was suitable for small target counting. The clear fry image was obtained by adjusting the Angle of camera and light source, and then image preprocessing methods such as background difference, Gaussian filter, global gray linear transformation and morphological processing were used to achieve the foreground segmentation and preliminary positioning of fry. In order to effectively unify the sample space and the target space to be identified, the minimum external moment is used to regularize the foreground image initially, and the regularized sample is used to build the image sample set. In the identification stage of turbot fry, the same pretreatment method was used to obtain the target region to be identified, and ResNet34 framework was introduced as the identification model to realize the accurate identification of the target region seed. Finally, the number of turbot seedlings was calculated by counting the number of targets to be identified. The experimental results showed that the method achieved good accuracy in the identification and counting of tiny fry. The average accuracy of the recognition of turbot fry using ResNet34 framework reached 94.27%, which was 8.64 percentage points higher than that of SVM method and AlexNet method, and 7.4 percentage points higher than that of SVM method. It is better than ResNet18 (identification accuracy 93.21%) and ResNet50 (identification accuracy 93.83%) and other similar structures. The average accuracy rate of fry counting of the model in this paper is 96.28%, indicating that the proposed sample set construction and identification method can meet the needs of small target counting, and can provide technical reference for fish fry counting.

Key words: