The work aims to propose a trademark image retrieval scheme with deep learning coupled sparse semantic measure to effectively restrain noise interference and reduce redundant feature dimension, with respect to the problem of semantic gap in trademark image retrieval. Firstly, according to the unsupervised learning mechanism composed of convolution and cistern, the multi-layer feature extraction of input trademark image were carried out to output the one-dimensional feature vector. Then, the L2-support vector machine (L2-SVM) was used to classify the feature vectors, and the multilevel features were obtained based on the training with feature vectors. Then, according to the multilevel feature of the trademark image and the heterogeneous data structure of the user's label information, a sparse semantic measure method was designed for similar retrieval, which reduced the semantic gap. In addition, a mixed norm was introduced as a sparse constraint of similarity measure to suppress the redundant feature dimension and noise in the original input space and optimize the retrieval results. The experiment showed that, compared with the current popular trademark retrieval scheme, the proposed algorithm had higher retrieval accuracy, whose output results only had one irrelevant image. The proposed scheme has higher retrieval precision and stronger robustness, and it has good application value in trademark detection, trademark protection and so on.