The body condition score （BCS） of dairy cows is one of the important indicators of animal health and welfare in precision animal husbandry farm，and an important basis for decision-making and management. Traditional methods of assessing body condition are mainly manual assessments. The traditional method is manual evaluation，which relies on human visual or tactile scoring evaluation of specific areas of the cow's body. Although the cost of manual method is low，it is time-consuming and labor-consuming. The manual evaluation has the disadvantages of subjectivity and low repeatability of scoring results. With the development of artificial intelligence，deep learning technology has been widely used in monitoring animal information. However，there is still a need for an efficient and real-time method of monitoring body condition of cow to meet the needs of commercialization. An improved lightweight attention mechanism network model （Shuffle-ECANet） was proposed to solve the problems mentioned above. Firstly，8 972 image samples containing the tail of cows were selected，and the body condition of cows was manually scored by animal husbandry experts to construct a relevant dataset. Then，an efficient channel attention module was introduced into the feature extraction structure of lightweight ShuffleNet-v2 to strengthen the network's ability to extract body condition features of cow. The H-Swish activation function was used to avoid neuronal necrosis. Finally，the Shuffle-ECANet network model was obtained by further simplifying the network structure. Three evaluation indicators including precision，recall，and F₁ were selected to evaluate the performance of models. Four models including EfficientNet-v1，MobileNet-v3，ShuffleNet-v2 1× and ResNet34 were used for comparative analysis to verify the performance of Shuffle-ECANet network model. The results showed that the Shuffle-ECANet model outperformed EfficientNet-v1，MobileNet-v3，ShuffleNet-v2 1× and ResNet34 in the results of evaluating body condition with BCS estimations within 0，0.25 and 0.50 units，respectively. The effectiveness of Shuffle-ECANet method was proved as well. The lightweight Shuffle-ECANet model proposed had an accuracy of more than 97% for each category，indicating that the model can distinguish different body conditions of cows effectively. It will provide the possibility for the refined management of individual body condition of dairy cows in large-scale pastures and a basis for the future application in low-computing power equipment，and a theoretical basis and idea for the commercialization of scoring body condition of cow.