Objective: To reduce the influence of data difference and spectral feature redundancy on beef quality recognition. Methods: A beef quality recognition method based on classification feature extraction and deep learning was proposed. The spectral feature extraction method of classified beef was designed, and the improved DPeak algorithm was used for adaptive clustering analysis of spectral data to realize the difference analysis of data. The objective function of beef spectral feature extraction was defined and solved by discrete lion swarm algorithm. The optimal spectral feature subset of each classification was extracted to minimize feature redundancy. The improved lion swarm algorithm (ILSO) was used to optimize the support vector machine (SVM) model parameters corresponding to each classification, and a beef quality recognition model integrating classification feature extraction and ILSO optimized SVM was proposed to complete the classification and evaluation of beef quality. Results: Compared with SSA-SVM and PCA-SVM, the recognition accuracy of this model is improved about 12.3%~14.5%. Conclusion: The beef quality recognition model based on classification feature extraction and deep learning can improve the accuracy of beef quality recognition.