Due to the low accuracy of block recognition in the process of feature extraction, traditional methods have poor extraction effect. In this context, deep reinforcement learning theory is introduced to carry out the extraction of visual communication image features. First, the image is divided into multiple rectangular areas, the maximum value of its internal elements is output, and the feature information output through convolution and pooling is integrated. Then, deep reinforcement learning is applied to maximize the accumulation of rewards for agents, and through interaction with the environment, a reward value as a training feedback signal is obtained. In order to reduce the amount of data and eliminate redundant data, different coding algorithms are used in the data compression process. At the same time, the LBP algorithm is used to extract robust information and correlate the weighted pattern pairs in the sliding window. Finally, a projected fuzzy clustering model is designed that can more effectively retain useful information in high-dimensional data. Using this model, the pixels in the image are iteratively divided into K clusters, and the features are extracted by calculating the Euclidean distance between the pixels and the cluster center. The test results show that using the proposed method, the block feature recognition rate can reach 95% and the feature extraction effect is good.