Reasearch And Application Of Electrical Equipment Instance Segmentation In RGB-T Images

When deep neural networks are utilized to segment images of electrical equipment,it is faced with the problems of expensive mask annotation and limited computing resources.Towards the demand of inspection robots and wearable operation assistant systems for electrical equipment image instance segmentation,the weakly-supervised and the semi-supervised methods,as well as the model compression approach,are presented.The electrical equipment instance segmentation module is developed and used in typical application scenarios such as insulator fault detection and power grid personnel knowledge training.The effectiveness of the model is verified by experiments.Dealing with the problem of expensive mask annotation in electrical equipment instance segmentation,the instance segmentation approach based on weak annotation is explored.First of all,a novel RGB-T automatic annotation method based on thermal image guidance is proposed,with image-level label supervision only.Moreover,to alleviate noise as well as to improve boundary segmentation,a pregressively optimized model(POM)is given,which ultilizes the fully-connected conditional filed(CRF)and the constrain-to-boundary loss to specify fine-detailed boundaries of each object.The proposed POM also exploits the self-paced learning technology to train the segmentation network from simple to complex in order to solve the oversimplification problem for training samples caused by the resolution difference between RGB-T images.For the case that only RGB images partially are annotated can be used,a semi-supervised instance segmentation based on progressively adversarial learning is proposed.Firstly,the parameter initialization of the backbone network of the instance segmentation model is carried out by using the deep convolutional generative adversarial network,and then the model and a discriminator are used to form the adversarial training,so as to realize the optimization of mash branch via the unlabeled data.Finally,the model re-optimization is realized based on the discriminator.Through the training above,the potential structure of unlabeled data is learned.To meet the needs of deploying the instance segmentation model on the systems with limited computing resources,such as inspection robots and wearable operation assistant devices,the model compression and acceleration method is explored.An effective instance segmentation model based on Mobile Net V3 and knowledge distillation is presented,and the model quantization is exploited to compress the model,which is mainly carried out by using a low-bit data format.Based on the above instance segmentation technology research,two typical applications of electrical equipment instance segmentation technology are constructed,which verifies its effectiveness.