To optimize security screening efficiency and reduce the risk of false detections and missed detections by security personnel, AI methods have been employed to detect prohibited items in X-ray images. However, the performance of such methods is hindered by the existence of small objects, as well as factors such as overlapping items, noise interference, and lack of detail. To improve detection accuracy, we propose a neural network model focused on small object detection in X-ray images, called Slice-based Recognition of Small Objects (SRSO). Firstly, to better detect small objects, we employ slicing aided inference and fine-tuning pipeline, dividing the X-ray image into overlapping blocks for detection, allowing small objects to have relatively larger pixel areas when input into the network. Secondly, to address issues of noise interference and lack of detail for small objects, we introduce a squeeze and excitation attention module into the network, which uses global information to selectively emphasize informative features and suppress less useful features. The experimental results on the SIXray dataset show that the SRSO method achieves satisfactory results in the precision, recall, f1 score, mAP, and AP(small) metrics. Importantly, the method is capable of maintaining real-time inference capability, demonstrating the effectiveness of our proposed SRSO method in improving X-ray security inspection models. The model and source code will be released upon acceptance.