An evaluation method for cooked fish fiber degree was constructed based on the convolutional neural network in order to break the situation that there was no systematic evaluation method for cooked fish fiber taste. A dataset was established by collecting microscopic images of cooked fish samples with different fiber degrees. The dataset was randomly divided into training dataset and testing dataset in the ratio of 8∶2. The constructed models were trained on the training dataset and their recognition performance were evaluated by the testing dataset. The result showed that the 34-layer ResNet model win in convergence speed and accuracy when choosing network depth. Four ResNet models with different depths were better than AlexNet， VGG-16 and GoogLeNet models in the best recognition accuracy. The average accuracy， precision， sensitivity and specificity and AUC of ResNet-34 were 96.94%， 91.26%， 91.00%， 98.13% and 99.19%， respectively， which proved that the evaluation method based on ResNet-34 model could objectively and accurately identify the degree of cooked fish fiber.