The service robot has been developed rapidly in recent years, its application algorithms are constantly alternating, and the item detection algorithm is one of them. Under the premise of ensuring item recognition accuracy, the item detection speed determines the efficiency of robot item capture. Therefore, this paper will take the long distance and small item scene as the test scene, and improve the existing network model. The aim is to improve the detection speed on the premise of ensuring detection accuracy. Mask regionsbased convolution neural network (Mask RCNN) is a widely used algorithm in the field of item detection. Through the studying its network structure, it is found that the mask branch and excessive full connection layers will take up a lot of network detection time. At the same time, the feature map extracted with mask RCNN has a higher dimension, which will take up a lot of computing memory and produce a large number of computing tasks. To tackle these problems, in this paper, the mask RCNN network is improved by removing the mask branch and redundant full link layer, the lighthead RCNN (LHRCNN) is introduced into the mask RCNN network, and the anchor ratio in the region proposal network (RPN) is adjusted. Finally, the improved Mask RCNN network was tested on the home service robot platform with Kinect Ⅱ. The test results demonstrate that compared with the original mask RCNN, the improved Mask RCNN network can greatly improve the running speed of the algorithm, while ensure the detection accuracy at the same time. The detection time is shortened by more than two times, and the proposed method improves the efficiency of the item catch task of service robot.