During the production，operation and processing of duck egg，the egg shells are easily broken and microorganisms including bacteria tend to invade the egg from the shell cracks，which in turn affect the quality of the eggs and damage economic benefits of production. An identification algorithm of duck-egg shell crack based on MEL spectrum was established by using ResNet34 network model to solve the problem of the subjectivity and large fluctuation of accuracy in manual identification of duck egg shell cracks. First， the egg knocker was used to collect the sound data， and the audio was transformed into the MEL spectrum graph to construct the dataset of MEL spectrum graph. Then the ResNet34 model was built and the transfer learning mechanism was introduced to train the model. The gradient was updated by Adam optimization algorithm， the attention mechanism module was added， and the convolution structure was replaced by a deeply separable convolution to improve the network model. The parameters were adjusted for optimization and the duck egg shell cracks were identified with the model.The results showed that the average detection accuracy of the ResNet34DP_CA enhanced network model was 92.4%，which was 5.5 percentage points higher than that of the original ResNet34 network model. The quantity of parameter was reduced by 32%. Compared with other network models including VGG16， MobileNetv2 and EfficientNet， the average accuracy was improved by 10.9， 13.7 and 16.3 percentage points，respectively. The recognition time was 21.5 ms. It is indicated that the established identification algorithm of duck-egg shell crack based on Mel spectrogram and the improved ResNet34 model can efficiently identify the duck-egg shell cracks. It will be of great significance to improve the economic benefits of production and to build an intelligent and modern poultry factory.