Deep Learning Based Model Compression And Acceleration In Foreign Object Detection On Transmission Lines

In recent years,with the excellent performance of Convolutional Neural Network(CNN)in the field of computer vision,more and more researchers are applying it in the field of power inspection.However,the increasing depth and parameters of deep learning models lead to very large storage and computational resources required for model deployment,making it difficult to deploy in resource-constrained embedded devices.Based on this,for the problems of large size,many parameters,and difficult deployment of deep learning-based foreign object classification and detection models,the main work in this paper is carried out as follows:1.To address the problem of the large size of deep learning-based foreign object classification models,this paper proposes a quantization-based improved the Goog Le Net foreign object classification model.First,the performance of several classical CNN models for foreign object classification on transmission lines is evaluated.Then,the Goog Le Net with the best classification performance is improved based on small target detection,inception module,and detection class compression,and the results show that the improved model improves the validation accuracy by 4.1% and compresses the size by 15.172%.Finally,the model weights were quantified and studied.The experimental results show that the FP16 weighting model reduces the volume by half without losing accuracy,which proves the effectiveness of the quantization strategy.2.To address the problems that deep learning-based foreign object classification models are computationally intensive and difficult to deploy on the embedded device,this paper proposes a pruning-based improved the Mobile Net V2 lightweight foreign object classification model.First,the performance of several lightweight models designed for mobile device is evaluated for foreign object classification on transmission lines.Then,a pruning-based model compression and acceleration study is conducted on Mobile Net V2,which has the best classification performance,and the optimal pruning ratio is determined.The experimental results show that the pruned model compresses the volume by 51.77%and improves the inference speed by 2.5 times when the verification accuracy loses only1.14%,which proves the effectiveness of the pruning strategy.Meanwhile,the pruned Mobile Net V2 model is smaller in size and faster in inference speed than the Goog Le Net foreign object classification model with FP16 weights.3.In the actual foreign object inspection process,it is not only necessary to classify the foreign objects but also sometimes it is necessary to frame the detected foreign objects.Based on this,this paper then researches the foreign object detection model based on YOLOv3.At the same time,this paper prunes and quantifies the YOLOv3 foreign object detection model to address the problems of large size,computation,and deployment difficulties of the YOLOv3 foreign object detection model.First,the feature extraction network of YOLOv3 is replaced by Darknet53 with the lightweight Mobile Net V1,which reduces the model size and computation.Then,the model compression and acceleration are studied by pruning and then quantization.The experimental results show that the model size after pruning and quantization is compressed by 79.661%,the inference speed is improved by 1.6 times,and the average precision is improved to 0.8891.While ensuring the model accuracy,the model becomes lightweight and consumes less hardware storage and computing resources.Finally,in the deployment test on Raspberry Pi,the model inference speed is only 190.72 ms,which can achieve real-time detection.