Underwater acoustic target recognition is to classify the targets through collecting the signals of the underwater acoustic targets and has very important and extensive applications in the fields of ocean exploration and monitoring technology. Due to the complexity of the marine environment, the diversity of target ship engines, and the background noise, underwater acoustic target recognition is difficult. Traditional feature extraction methods cannot extract effective feature representations to fully represent the targets. In order to solve this problem, we propose an underwater acoustic target recognition algorithm based on the improved bag of visual words. Specifically, this algorithm adopts the bag of visual words to extract high-dimensional features in a spectro gram and then adjusts the weights of the obtained feature vectors using the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm commonly used in the field of natural language processing. Furthermore, the vectors are input to a Multi Layer Perceptron (MLP) to classify and recognize the underwater acoustic targets. The experimental results show that the recognition algorithm proposed in this study achieves an accuracy of 92.53%, which is a significant improvement in comparison with the best Deep Boltzmann Machine (DBM) algorithm at present.