In order to detect underwater crabs, a detection method based on machine vision was proposed, which can produce necessary feedback data on the number and distribution of crabs to automatic bait casting boat in real time so that the boat can cast baits precisely. An underwater camera with LED light installed at the bottom of the boat was used to capture images of crabs. These images taken under water were enhanced by optimized retinex filter firstly, which can make images clearer and enhance the details in images. Then, a dataset included original images captured underwater, captured from a laboratory and downloaded from webs was built. There were totally 3500 labelled original images in the dataset. The dataset was augmented and divided into training and test datasets. Finally, a deep convolution neural network, YOLO V3 was trained to detect the crabs by training dataset. The mean average precision of trained network reached 86.42% on test dataset, the detection precision for underwater crabs was 96.65% and the recall ratio was 91.30%. Compared with crabs with big size, the crabs with small size were more difficult to be detected. Compared with other methods for object detection, YOLO V3 can reach a high level of both recognition precision and speed. The recognition speed of the proposed method was 10.67f/s, which was higher than that of other methods on the same hardware platform. Therefore, the proposed method was real time and had application values.