Abstract: ‍ ‍Synthetic aperture radar （SAR） possesses all-weather and all-day imaging capabilities， which are highly important for SAR image target detection for military and civilian applications. In SAR image target detection， repeated detection occurs due to complex backgrounds during imaging and interference from non-target objects. Traditional deep-learning networks used for SAR image detection reduce the probability of repeated detection by increasing feature extraction networks， non-maximum suppression， and other processes. However， improper threshold settings and overlapping detection targets can still lead to false alarms and missed detections. To address this， this paper introduces a transformer-based target detection model with binary matching loss. Compared to traditional SAR image detection networks， binary matching utilizes the Hungarian algorithm to perform one-to-one matching between predicted boxes and candidate boxes， thereby identifying the best matching pairs and avoiding repeated detection of the same target. During matching， redundant candidate boxes are automatically ignored and classified as background， eliminating false alarms caused by repeated detection and omitting the need for non-maximum suppression operations. Moreover， the matching results can directly affect the model’s output， achieving end-to-end detection optimization and transforming the target detection task into a set prediction problem. Through a fixed set of learnable position encodings， effective associations between targets and image features are established without relying on prior knowledge or preprocessing steps， greatly simplifying the training and deployment processes compared to traditional methods. To evaluate the effectiveness and reliability of the model， comparisons were made with current state-of-the-art target detection models on the SAR-AIRcraft-1.0 dataset， achieving good detection accuracy， while ensuring a high recall rate， demonstrating the superior performance of the model.