Abstract: The detection effect of infrared image insulator strings is affected by the environment in the power grid inspection. The combination of saliency detection and improved convolution neural network (Fast-CNN network) is proposed for insulator feature detection. Firstly, superpixels are used to describe the overall information of each region, saliency features are calculated based on the characteristic covariance information of each superpixel. Then the salient features are extracted by regional modular extraction and local complexity contrast. At the same time, the salient features extracted from the two methods are respectively input into the improved Fast-CNN network for salient region detection, a dynamic adaptive pool model is proposed, the cosine window is introduced to deal with the middle layer. Finally, the characteristics of insulators are obtained through iterative training. It can avoid full graph search for the CNN model. The proposed algorithm is tested in the infrared image library， the F-Measure and the average error MAE of the proposed algorithm are better than the current popular algorithms.

Key words:  , machine vision； deep learning； salient compute；insulator detection； Fast-CNN