Accurate and stable image feature extraction is of great significance to computer vision applications such as image stitching, 3D reconstruction, and feature-based visual simultaneous localization and mapping (SLAM). In the nuclear radiation environment, the captured images have the problems such as many noise points, large noise blocks and the noises are easy to identify as features through the traditional feature extraction methods. An against nuclear feature (ANF) extraction algorithm is proposed based on the noise distribution characteristics of the γ rays affected images which is collected by the source blockage failure of an irradiation factory. Firstly, the red, green, blue (RGB) characteristics and grayscale characteristics of each pixel in the image under nuclear radiation are analyzed to obtain the pixels which are suspected as noise points. Then, the features are extracted by the traditional feature extraction algorithm. Finally, the Euclidean distances between the features and the suspected noise points are used to sort and filter the features, and the features which are suspected as noises with high probability are eliminated. The experiments based on the standard image data set combined noises and collected images in the real nuclear radiation environment show that the ANF method is more stable than the traditional features from accelerated segment test (FAST) method and binary robust invariant scalable keypoints (BRISK) method in extracting features, and can improve the effect of feature extraction and reduce the matching error rate.