Aiming at the difficulty of identifying multi-target prohibited item in X-ray security images, a multi-target prohibited item recognition algorithm is proposed in this paper. First, considering practical application requirements, network performance and running speed, the residual network (ResNet50) is used as the backbone network, and a local reinforcement module is added to compensate for the checkerboard phenomenon caused by dilate convolution. Then, the features of different levels are processed by the dilated residual feature enhancement module and the transformable dilated space pyramid pooling respectively, and the multi-scale characteristics of prohibited item are adaptively learned. Finally, the attention mechanism is introduced to strengthen the learning ability of key channels and realize the feature focusing in the spatial dimension, so as to strengthen the detailed representation ability of the prohibited item area. The test results on the security inspection prohibited item image dataset show that compared with other comparison algorithms, the algorithm can achieve better segmentation accuracy on the premise of ensuring real-time performance, the mean intersection-over-union is 82.26%, and the image processing speed is 16.21 frame/s.