Research On Autonomous-learning Image Semantic Segmentation Algorithm And Application In Insulator Damage Detection

The image semantic segmentation technology is one of the most important topics in the field of image processing,fast and accurate semantic segmentation has always been the research goal of the researchers.With the development of artificial intelligence,image semantic segmentation algorithms are developed gradually based on deep learning methods.However,pixels are the basic unit in the processing of current image semantic segmentation algorithms that based on deep learning,and there is a lot of redundancy,which increases the time complexity and space complexity of the algorithm.Moreover,the accurate segmentation results often depend on a more complex network structure,and a large number of parameters make network learning difficult.In order to make the image semantic segmentation algorithm more accurate and efficient,this paper proposed an Autonomous-learning image semantic segmentation algorithm and applied it to insulator damage detection.The main work of this paper can be summarized as follows:(1)It proposed ALIC(Autonomous-attachment Linear Iterative Clustering)superpixel segmentation algorithm.It speeds up the algorithm by using a simple and efficient method of distance measurement.At the same time,the natural continuity between pixels is fully considered by the allocation principle of Autonomous-attachment,each pixel can share the label with its neighbors,which improved the accuracy of superpixel segmentation.The experimental results proved that the method in this paper can achieve convergence within 5 iterations,and reducing the calculation time while ensuring higher segmentation accuracy.(2)It proposed an improved FCN(Fully Convolutional Networks)learning neural network.This learning network is using VGG-16 Net as the backbone network and optimized it with a deconvolutional layer instead of three fully connected layers,which,simplify the network structure.Meanwhile,it upsampling the output of each convolution block in the network and get a high-resolution feature map.In the end,it concat the feature map and get the final output result.The improved FCN solved the problem of low semantic segmentation accuracy caused by the loss of detailed features extracted by shallow networks in FCN.(3)This paper combines ALIC superpixel segmentation algorithm and improved FCN to form an Autonomous-learning image semantic segmentation algorithm.The image upgraded to superpixel image by Autonomous-attachment linear iterative clustering and obtains the result of semantic segmentation by improved FCN.In the experiment of public database from PASCAL VOC2012,the method in this paper obtained a MIoU of 67.3% and a Recall rate of 80.5%,more importantly,the training time and operation time are less than the common FCN-based image semantic segmentation algorithm.(4)This paper use insulator as a data set and finished a simulation experiment of insulator damage detection on Autonomous-learning image semantic segmentation algorithm,61.1% of MIoU and a Recall rate of 75.8% were obtained in this experiment.By analyzing the experimental results,it can be concluded that the Autonomous-learning image semantic segmentation algorithm is feasible in the application of insulator damage detection.At last,this paper made a prospect for the further application of insulator damage detection.