In order to eliminate the influence of background clutter and noise in the infrared images for the detection of infrared dim targets at sea, an infrared image dim target detection method based on temporal-spatial non-local similarity is proposed. The method makes full use of the non-local auto-correlation of the sea surface background image patch between the infrared image sequences of adjacent frames and the similarity between non-local background image patch in the frame. Based on this characteristic, a temporal-spatial image patch model is introduced. The model is effectively solved by accelerating the proximal gradient method. The experimental results show that compared with the traditional infrared dim target detection method, the proposed method can not only preserve the feature of the target more effectively but also increase the peak signal to noise ratio of the detected infrared image by 1.2 times and improve the signal to clutter ratio by 1.8 times.