Estimation Method of Fry Body Length Based on Visible Spectrum
Li Zhen-bo,Niu Bing-shan,Peng Fang,Li Guang-yao
DOI: https://doi.org/10.3964/j.issn.1000-0593(2020)04-1243-08
2020-01-01
Spectroscopy and spectral analysis
Abstract:In the process of fry breeding, the same breeding pond will appear that the individual large fry attacks small individual fry, and the individual small fry will suffer injury or even death, resulting in economic loss, and the fry pond and the selling price are mainly related to their body length parameters, so separation of different sizes of fry is required. The classification of fry is mainly dependent on mesh screens of different sizes, which is time consuming and laborious, and is easy to cause damage to fry. Aiming at the low efficiency and lack of scientific guidance of traditional artificial separation methods, this paper proposes a study on the estimation method of fry body length based on visible spectrum, which can calculate the length and classify the fry according to the fry image. In order to obtain the body length of the fry accurately and without loss, a method for estimating the length of the fry based on the migration learning ResNet50 model was proposed. Firstly, images of different lengths of fry taken under the same height conditions were collected. At the same time, the actual length of the fry was manually measured as the label of the data set. The four migration learning models AlexNet, VGG16, GoogLeNet, ResNet50 were used to estimate the body length of the fry, and the verification set was passed. Accuracy, test set accuracy, and running time of different methods were analyzed. The accuracy rate of AlexNet model verification set was 90. 04% , the test set accuracy rate was 89. 82% , and the running time was 52 minutes and 3 seconds; the VGG16 model verification set accuracy rate was 91. 01% , the test set accuracy rate was 91. 17% , and the running time was 131 minutes and 37 seconds; the accuracy rate of GoogLeNet model verification set was 88. 02% , the test set accuracy rate is 88. 39% , and the running time was 45 minutes and 2 seconds; the ResNet50 model verification set accuracy rate was 91. 92% , the test set accuracy rate was 91. 09% , and the running time was 99 minutes and 17 seconds; then determined ResNet50. The model had a 50-layer Residual Network architecture. The migration learning method was used to transfer the parameters of the convolution layer trained on ImageNet to the model used in this training, and the softmax layer was adjusted to adapt to this problem. The experimental results on the fry datasets of 6677 samples from 10 different lengths showed that the method can be effectively used for fry classification, and the model was optimized by stratifying the number of layers, the number of iterations, the learning rate of the model ResNet50 and the Mini Batch Size. The experimental results showed that when the migration learning model had 30 migration layers, the number of iterations was 6, the learning rate was 0. 001, and the Mini Batch Size was 10, so the method effect was optimal. The accuracy of the model verification set was 94. 31% , and the accuracy of the test set was 93. 93%. The algorithm in this paper estimates the accuracy of fry body length by about 2% compared with the traditional image processing method. In the future, in the actual production scenario, the method can be nested into the fry body length separation device, the scientific research can be put into actual production, the fry damage can be reduced, and the foundation for the future unmanned fishery can be laid.