Research On Key Technologies Of Image Target Detection Based On Deep Learning

Intelligent target detection is widely used in the fields of robot navigation,intelligent video surveillance,and industrial inspection.In order to improve the accuracy and precision of intelligent target detection,the thesis research on three key technologies,including image denoising algorithm,target detection algorithm based on candidate regions and image semantic segmentation.The main research contents are as follows:In the existing noise reduction algorithm,a large amount of high frequency information would be lost in the denoising process,which causes the image to be too smooth and obscure after noise reduction.Therefore,a composite noise reduction network composed of the convolutional autoencoder network and the feature reconstruction network is proposed.In the composite noise reduction network,the cross-connected structure is used to fuse feature information form the convolutional autoencoder network into the feature reconstruction network for feature compensation,and the step-by-step training method is used,parameters are optimized separately and finally all parameters are optimized again.Experimental results show that the proposed method has better noise reduction performance for both Gaussian noise and salt and pepper noise.The border regression is a key technique of the regional convolution neural network to locate the target.However,it relies on the border label information of a large number of sample data.Therefore,it is inefficient to generate the training sample set,and the location of the target is also inaccurate.For this,a novel target detection method based on the CNN and the particle search is proposed.A small number of probe particles are generated to roughly locate the target.The CNN is used to extract the image features,determine the target probability,and recognize the category of the target.A large number of searching particles are placed near the region where the target features are detected by the probe particles.The nearest neighbor clustering algorithm is used to classify the particles,which are recognized as the same divide into different target sets.The positions of the targets can be determined by the bounding rectangle of the searching particles in the same target set.The simulation results show that the correctness of the recognition can be slightlyimproved,and the accuracy of the location can be significantly improved.In the full convolutional network and U-shaped convolutional network,down sampling will cause the loss of image characteristic information and the accuracy of image semantic segmentation will be affected.For this,a convolutional network model for image semantic segmentation is designed to improve the accuracy of segmentation.In the model,the U-shaped network structure is used.In order to avoid network degradation problems,the residual structure is introduced.A multi-scale pooled structure is designed for feature compensation.The regularization layer is added to prevent over-fitting and speed up network training.The conditional random field is used for post-correction to improve the accuracy of image semantic segmentation.The experimental results show that the image segmentation method has better segmentation effect.The dissertation completed the construction of the robotic arm sorting system based on the target detection algorithm,the image denoising algorithm and image semantic segmentation algorithm are applied to the system and experiments were performed to verify the effectiveness of the proposed algorithm.