Research On Damaged Video Frame Detection And Optimization Algorithm Based On Convolutional Neural Network

The fierce market competition in the online video industry makes it important to maintain good video quality,so effectively detecting damaged video frames has become an important issue.The traditional damaged video frame detection algorithm is directly affected by the edge features of the mosaic,which is highly sensitive to the scene so as to have low detection accuracy.In recent years,convolutional neural network(CNN)has been widely used in image processing and has performed well.However,CNN has a huge amount of computation and parameters.Therefore,this paper applies it to the detection of damaged video frames,and designs network acceleration and network compression algorithms to further optimize the network as well as improve the detection accuracy.The research contents of this paper are as follows:1.To solve the problem that the traditional damaged video frame detection algorithm is not so accurate,we construct a small-scale detection model based on atrous convolution,residual learning and dense connection techniques of CNN.The experimental results show that this detection model can adapt well to different video scenes with high accuracy and stability.2.To meet the real-time requirement of video frame detection,by analyzing the time-consuming reasons in convolutional operation,this paper designs a network acceleration structure Group-Inception based on feature map channel decomposition and kernel channel decomposition.The results show that this structure can greatly reduce the computational complexity while ensuring good detection effect.3.To meet the low resource consumption requirements,by analyzing the source of the parameters in the CNN,this paper proposes a new network pruning algorithm to compress the network.Firstly,through the analysis of network weights and data propagation process,a node importance evaluation algorithm based on data flow is proposed.Then,the memory mechanism is introduced in order to avoid false pruning due to fewer evaluation samples.Finally,a soft pruning strategy based on importance sampling is proposed to improve the stability of pruning.Combining with the above researches,a network optimization closed-loop for network acceleration and compression is constructed,and then the final model is obtained through iterative training.Compared with the traditional network,this optimized model has obvious effect of acceleration and compression while ensuring the detection accuracy.Therefore,it can be well applied to video detection scenarioswith strict requirements on accuracy,real-time performance and resource consumption.At the same time,this network optimization algorithm can be applied to other CNN scenarios,which has great research and application value.