To increase the defect detection accuracy on substation equipment and thus ensure the operation safety of the substation, this study proposes a defect detection algorithm based on an improved YOLOv4. Unlike the original YOLOv4, the new algorithm replaces the fully connected layers with one-dimensional convolution to optimize the convolutional block attention module (CBAM), which is then embedded into the backbone network to enhance the feature extraction ability. Meanwhile, dilated convolution is used in feature fusion layers for expanding the receptive field and aggregating broader semantic information. The algorithm is tested on images captured in real substation scenes and achieves a mean average precision (mAP) of 86.97%, an increase of 2.78% on that of the original YOLOv4. Experimental results show that the proposed algorithm can improve the network performance and is thus more suitable for defect detection on substation equipment.