基于迁移学习和金字塔卷积网络的河蟹个体图像识别方法研究

针对目前河蟹追溯成本高、消费者无法细粒度地追溯单体河蟹信息等问题,提出一种基于迁移学习和金字塔卷积的河蟹背甲图像个体识别算法。该算法使用金字塔卷积层替换普通残差卷积块构建网络模型,可以从蟹背图像中提取多尺度、深层次的特征信息。结果显示:采用金字塔卷积结构的Resnet34和Resnet50的准确率分别为98.38%、98.51%,与使用普通卷积层的模型相比,准确率提升5.49%、1.3%,而当模型深度达到101层时,模型性能不再明显提升。与使用金字塔卷积结构的全新学习模型相比,迁移学习方法的训练收敛迭代轮次从20轮降低至5轮,此时模型准确率为98.88%,较全新学习的准确率提升0.37%,同时弥补了样本量较少的问题。该研究为河蟹个体识别追溯提供了理论依据和技术支持。

Research on individual image recognition of river crab based ontransfer learning and pyramid convolution network

Aiming at the problems of high tracing cost of river crab and consumers'inability to trace monomer river crab information in fine grain,an individual recognition algorithm of river crab carapace image based on transfer learning and pyramid convolution is proposed.The algorithm uses pyramid convolution layer to replace ordinary residual convolution block to build a network model,which can extract multi-scale and deep-seated feature information from crab back image.The results show that the accuracy of resnet34 and resnet50 with pyramid convolution structure are 98.38%and 98.51%respectively.Compared with the model with ordinary convolution layer,the accuracy is improved by 5.49%and 1.3%.When the depth of the model reaches 101 layers,the performance of the model is no longer significantly improved.Compared with the new learning model using pyramid convolution structure,the training convergence iteration rounds of the transfer learning method are reduced from 20 rounds to 5 rounds.At this time,the accuracy of the model is 98.88%,which is 0.37%higher than that of the new learning.At the same time,it makes up for the problem of small sample size.This study provides theoretical basis and technical support for individual identification and traceability of river crab.