Beauty Pageant of Koi Method Based on Improved ResNeXt50 Residual Network

Abstract:

There is a high similarity among different levels of the koi for beauty pageant, and beauty grading for koi is currently done manually. To solve these problems of low efficiency, strong subjectivity and high cost of manual beauty pageants, a sorting method for koi beauty pageant was proposed based on transfer learning and improved ResNeXt50 residual network. Firstly, a rank dataset was constructed for the beauty pageant on Kohaku, Taisho and Showa koi. Secondly, the transfer learning strategy was adopted to improve the training speed and improve the ResNeXt50 model from three aspects of SE attention module, Hardswish activation function and Ranger optimizer, further a SH-ResNeXt50 classification model was proposed and constructed for koi pageant. The experimental results showed that the SH-ResNeXt50 model effectively improved the sorting ability for koi beauty pageant, with an accuracy of 95.6% and a loss value of only 0.074, which was better than the commonly used AlexNet, GoogLeNet, ResNet50 and ResNeXt50 network models. Finally, the interpretability of SH-ResNeXt50 model was analyzed by Grad-CAM, and the results showed that the regions of interest of SHResNeXt50 model was basically consistent with those recognized by the humans. The approach proposed realized the intelligent sorting of different levels of koi beauty pageant with high similarity, which had reference significance for other biological level sorting with high similarity.