基于改进Deeplabv3+的大口黑鲈表型数据测量与品质预测方法

Phenotype Data Measurement and Quality Assessment of Largemouth Bass (Micropterus salmoides) Based on Improved Deeplabv3+

摘要:

大口黑鲈的肥满度、体质量及尺寸等是评估其品质的重要指标，针对上述数据手工测量操作繁琐、效率低下及关键点表型数据测量方法面积要素缺失等问题，本文提出一种语义分割模型结合最小外接轴对齐矩形的表型数据测量方法，并基于表型数据测量与计算结果完成大口黑鲈品质预测。首先通过使用CBAM(Convolutional block attention module)和SENet (Squeeze-and-excitation network)对Deeplabv3+模型进行改进，实现对大口黑鲈头部、躯干、尾部、鱼鳍等部位的高精度分割，然后使用最小外接轴对齐矩形完成大口黑鲈各部位长、高测量，通过各部位像素与矩形像素的比例完成面积测量；最后，基于测量结果完成体质量回归预测与肥满度计算，以实现大口黑鲈品质预测。结果表明，语义分割模型整体mIoU (Mean intersection over union)达到90.15%，在忽略鱼鳍影响后，mIoU达到94.02%，测量所得全长、体长、体高平均相对误差低于2.5%，头长、头高平均相对误差低于3.5%，面积测量误差低于4.5%。多项式体质量回归预测模型对体质量预测值与实际值的决定系数为0.97，平均相对误差低于4%，基于测量值的3种肥满度状态指数计算结果均接近实际值。该方法可以高效、准确地获取大口黑鲈的表型数据，并为进一步衡量鱼类生长状况与健康状况研究提供参考。

Abstract:

Phenotype data, weight, and condition factor of largemouth bass served as crucial basic information, providing a direct insight into the fish’s growth and health status in fishery aquaculture. In response to the problems of cumbersome and inefficient manual measurement of the above data, and the lack of area elements in phenotype data measurement methods based on key point, a phenotype data measurement method was proposed based on Deeplabv3+. And quality assessment was completed based on phenotype data measurement results. First of all, based on the analysis of the morphological characteristics of visible parts of the largemouth bass, the fish was divided into four parts: head, trunk, fins, and tail. Each of these parts was manually annotated with semantic labels by using Labelme software. Different batches images of largemouth bass were used as the dataset, and following the process of data enhancement, the dataset reached 1095 pieces, and the ratio of training set to validation set was 9∶1. Secondly, using convolutional block attention module (CBAM) and squeeze-and-excitation network (SENet) to improve the Deeplabv3+, the CBAM module adjusted feature map weights adaptively by utilizing channel attention and spatial attention, enabled the network to concentrate on the morphological features of largemouth bass, thereby enhanced segmentation accuracy. The SENet module mitigated channel redundancy in Deeplabv3+ network feature maps, thereby enhanced both parameter and computational efficiency. With the help of the above modules, the overall model could achieve high-precision segmentation of the head, trunk, tail, and fins of largemouth bass. Subsequently, the segmentation results were used to measure the length, height of each part by using the minimum axis-aligned bounding box. And the area was estimated based on the ratio of pixels in each part to box pixels. The area was fitted by using actual measured phenotype data, and the results were used as standard values to evaluate the accuracy of area estimation. Three commonly used fitting models were used to fit the weight of largemouth bass based on measurement results, and the best fitting model could be found by comparing the fitting results. Finally, the condition factors were computed by using body length and weight, and compared with actual condition factors to further validate the accuracy of the measurement data. The experimental result showed that the overall mIoU of the semantic segmentation model reached 90.15%, and after ignoring the influence of fish fins, the mIoU reached 94.02%. The mean relative errors of the measured total length (TL), body length (BL), body height (BH) , head length (HL) and head height (HH) were less than 3.5%. The mean relative error of area estimation was less than 4.5%. The correlation coefficient between the predicted and actual weight values by using polynomial models was 0.97, with the mean relative error less than 4%. The calculated results of the three condition factors based on the measured values were close to the actual values. This method could efficiently and accurately obtain phenotype data of largemouth bass and predict its growth status, providing a reference for the study of fish growth and health.