Abstract: Aiming at the low accuracy of nuclear waste detection under radiation environment, this paper proposes a nuclear waste detection algorithm named Dense-Dilated-YOLO V3 based on deep learning convolution neural network. The experimental results show that Dense-Dilated-YOLO V3 increases the network receptive field without increasing the parameters, effectively avoids the loss of image information, extracts more detailed features of the target in the radiation environment, and accurately detects the target under radiation environment. The rate reached 93.29%, which was 5.53% higher than the original algorithm, and the recall rate reached 91.73%, with an increase of 8.28%. It solved the problem of low accuracy of nuclear waste detection under complex radiation environment, and has better detection effect. It provides a new approach for nuclear waste detection.