With the increase of register trademark, there will consume many time and effort for human to retrieve and compare a trademark with large trademark database. In this paper, we dedicate to studying and developing an efficiency and precision system in this large trademark image database, which reduce the limitation of-traditional keyword annotation retrieval method and improve trademark register quality and speed. Retrieval efficiency and accuracy are two important issues in content-based image retrieval field. Content-based image retrieval (CBIR) has been extensively studied for many years and a number of techniques have been proposed. However, it is still a difficult task and the result is still far from practical. In this thesis, on the basis of generic CBIR theory and considering trademark image features, we propose a new retrieval system. This system achieves both desired efficiency and accuracy. The main contents of the thesis are as follows. 1. In preprocess stage, considering the difference between trademark image and general natural image, the trademark images are converted to binary images firstly. And then, removing noise, eliminating words and texts, normalization and edge detection are processed orderly which will eliminate the effect of transition and scale and increase the robust of feature extraction. 2. Multi-feature extraction method is adopted to improve the accuracy of system. We not only use some global feature such as invariant moment, edge direction, wavelet energy etc, but also introduce excellent local shape feature in character recognition field such as direction element feature (DEF), which emphasize local information of trademark. Lastly, we use probability principal component analysis (PPCA) to reduce dimension of original feature, which accelerate feature match. 3. Hierarchical retrieval and relevance feedback. Using hierarchical retrieval method, the accuracy and speed of retrieval are both ensured. Relevance feedback introduces subjective perception of user and improves the accuracy of retrieval The experiment results verify the valid of system. Though the system is designed for trademark image, most of the ideas and algorithms can be applied to other large image database retrieval problems. We believe that it is a difficult and challenge to develop an automatic retrieval method which models the human perception well. However, considering the time and effort for humans to perform such a retrieval task, an automatic retrieval method can provide a powerful tool to facilitate the retrieval task for human beings.