| Convolutional Neural Network(CNN)-based deep learning techniques have received a lot of attention and application in the field of computer vision in recent years due to the rapid advancement of artificial intelligence.Target detection is one of them and has improved detection results with the thorough application of convolutional neural network technology.Target detection is the fundamental technology of traditional computer vision applications such as face recognition,automatic driving,and medical imaging.The convolutional neural network model needs a more sophisticated network structure,a greater parameter scale,and a significant amount of calculation in order to achieve improved detection accuracy.These constraints limit the implementation of the deep neural network model on terminal devices with limited resources,impeding the advancement of neural networks.This research focuses on two areas of model optimization and calculation optimization in order to accelerate convolutional neural networks.In order to maximize the effectiveness of the model,two distinct model lightweight methods are proposed from the perspectives of pruning and quantization.At the same time,experiments are used to further confirm the correction.The magnitude of the model’s parameters and its computation volume are both greatly reduced by this approach.The detection job of video data is studied for actual application scenarios in this work,and a reasoning approach based on regions of interest is proposed.This method speeds up inference by decreasing the calculation of reasoning by reducing the data scale of the task processing.This paper’s precise work material is as follows:(1)This work suggests a structured pruning strategy based on mixed measures in an effort to address the one-sidedness issue with employing a single metric in structured pruning.In order to evaluate the significance of the convolution kernel and carry out pruning,this technique combines two parameter metrics,the sum of the absolute value of the convolution kernel weight and the L1 norm and the scaling factor,and uses the final mixed parameter standard MS(Mixed standard).In parallel,research on the neural network of the YOLO series was conducted.The findings indicated that employing mixed measures had a greater pruning impact than using a single metric.On the basis of ensuring the model’s accuracy,the pruning rate for YOLOv3 reached 77.45%,and for YOLOv3-tiny,it reached 82.60%.(2)This work suggests a quantization system based on mixed precision as an alternative to the conventional single-precision quantization.This method quantizes each convolutional layer separately to assess the quantization sensitivity of the convolutional layer,calculates the mean-square error(MSE)between the output of the quantized network and the original network,and then establishes the quantization sensitivity of each convolutional layer.This paper also explores the backbone component and the detection component of the network separately,analyzes the problem of "gradient disappearance" made worse by fine-tuning training after quantization,and presents the ultimate mixed precision quantization approach.The YOLO series network experiment findings show that using mixed precision quantization can increase compression rate with only a minor loss of precision.After quantization on the YOLOv3 network,the model size for one of them is decreased by 77.79%,and the precision m AP is equivalent to It is 0.92% less than the original network,and the accuracy m AP is 1.92% less on YOLOv3-tiny while the model size is 76.25%smaller on the latter.(3)In terms of practical application,this work suggests an inference technique based on regions of interest for the problem of target detection in video data.The algorithm separates the photos of specific consecutive frames into a group based on how similar the visuals of nearby frames in the movie are.Just the first image in a group of photographs is fully calculated when using forward reasoning.Then,when detecting the remaining images in a group,only calculate the data in the region of interest,and the data in the non-interest region,using the detection result of the first image to obtain the region of interest of the group and keeping the calculation result of the first image at the same time.Direct data copying from the first sheet’s calculation results is followed by the correction of the region of interest’s position.After experimental validation,the approach can successfully decrease the network’s computation load and speed up detection.The accuracy m AP of the YOLOv3 network only decreases by 0.5%,and the amount of calculation is decreased by 2.87 times,when the expansion ratio of the region of interest is set to 10% and the number of images in each group is 10 frames. |
