Study On The Security Object Detection And Recognition Based On Deep Learning

Convolutional neural network is one of the new kinds of neural network,which is combined with artificial neural network and deep learning technology.It is the first algorithm model which can train multi-layer network structure successfully.Today,convolution neural network has developed into an efficient method for object detection and recognition.The application of convolutional neural network to the detection of security objects in working site can realize the supervision of personal safety,equipment safety and construction safety,and it is an important dependence of safety assurance in workplace.Therefore,the object detection and recognition technology based on deep learning has important research significance and value.Under the application scenario of security objects detection in working site,this paper studies Faster RCNN models based on ResNet101.Firstly,this paper studies the activation function in convolutional neural network and proposes a new unsaturated activation function ReLU-Softplus.It combines the advantages of ReLU and Softplus activation functions,and corrects the distribution of input data in neural networks.ReLU-Softplus activation function not only solves the problem that the ReLU activation function easily make certain neuron parameters may never be activated and the corresponding parameters can never be updated,but also retains the advantage of fast convergence of convolution neural networks when use the ReLU activation function.Secondly,the traditional pooling method has some damage to the expression of global feature and the recognition accuracy of the model when extracting the feature.To solve this problem,the paper proposes the dynamic adaptive pooling algorithm based on the maximum pooling algorithm,which makes the convolutional neural network can extract more precise features when processing different pooled domains under different iterations.At the end of this paper,to solve the problem that detection of multiple small sample objects is prone to miscalculation,the optimization of the loss function is proposed.The optimization of the loss function mainly includes the use of the weight Softmax for the classification loss of the object,making the small sample object category more expensive to judge errors,in addition,making improvements to original Softmax losses by adding Center Loss item makes the trained network more cohesive,which enhances the robustness of the algorithm and reduces the occurrence of object misjudgment.This paper verifies the feasibility of the proposed algorithm on the experimental data set in practical application scenarios.Comparisons of the effects on the object classification and the position prediction are made between the improved model and Faster RCNN model.The final experimental results show that the improved model has higher recognition accuracy and more generalization,and the main advantage is the more precise effect of object position prediction,which reduces a lot of background and interference information,and makes the object position prediction more accurate.