Abstract: The inspection of typical fittings and their defects in transmission lines is a very important inspection task. Remarkable changes in the target size of fittings occur, and some of the fittings are small-scale targets, thus there exists the problem of low detection accuracy.Consequently, a detection method of typical fittings and its partial defects based on improved Cascade R-CNN is proposed. On the basis of the Cascade R-CNN model, the recursive feature pyramid structure is used for feature optimization, and hierarchical high-level semantic features are optimized vertically. In the horizontal direction, feedback connection structure gains backbone network characteristic map. At the same time, it is proposed to use NAS to obtain the atrous rate of the atrous convolution and to expand the receptive field, thus the multi-scale fittings features are extracted by the convolution more effectively. Experimental results prove that the proposed recursive feature pyramid is combined with the atrous convolution of the NAS (neural architecture search) search atrous rate to improve the method of Cascade R-CNN. To a certain extent, the proposed method can be adopted to solve the problem of low target detection accuracy of fittings. The performance index AP (average precision) value is increased by 6.72%, and the highest detection accuracy rate reaches 92.34%. The research has laid a good foundation for further fault diagnosis of typical fittings and realization of intelligent inspection.