Research On Feature Description Method Of Imige With Noise

At present,image description is a research hot topic in the field of computer vision and pattern recognition.It has achieved significant achievements in noiseless and low-noise images.In practical applications,when an image is acquired as an information carrier,the image will be introduced into the noise due to the influence of complex environment,image acquisition equipment,display equipment or other factors,which make it difficult to obtain features with high discrimination performance from the image with noise.Aiming at the drawbacks that traditional image description methods have poor robustness and insufficient ability of texture discrimination under noise conditions,this thesis proposes several optimized local feature extraction algorithms and feature selection algorithms.Moreover,extensive experiments are conducted on four public databases to verify the effectiveness of the proposed algorithms.The main contributions of this thesis are as follows:(1)In order to solve the problem that Local Difference Binary(LDB)operator can't recognize accurately when the texture of the central image block is similar to that of the surrounding image block,Multiple Extended Local Difference Binary(MELDB)is proposed.MELDB operator improves the LDB descriptor by increasing the pixel mean and gradient information in the diagonal and anti-diagonal directions.At the same time,each image is segmented several times with different sizes to extract features,and then the extracted features are cascaded.(2)In order to solve the problem that Genetic Algorithm(GA)and Binary Particle Swarm Optimization(BPSO)are easy to fall into local convergence,Fitness Diversity Synergy Genetic Algorithm(FDSGA)and Adaptive Learning Binary Particle Swarm Optimization(ALBPSO)are proposed.Specifically,FDSGA operator considers both fitness function and difference function,which makes the new optimal individual have a greater probability to enter the next iteration,while the value of learning factor and inertia weight in ALBPSO operator is self-adaptive learning according to the iteration process.(3)In order to solve the problem that the individual of FDSGA algorithm will be influenced by the assimilation and the individual of ALBPSO algorithm will be guided by the optimal individual,which makes the algorithm easy to fall into local optimum,IntelligentCooperative Feature Selection(ICFS)algorithm based on the cooperative evolution of FDSGA and ALBPSO is proposed.ICFS algorithm expands search space of the individual to find the more effective features of the individual by exchanging optimal individuals of the two algorithms in a certain round of iteration.(4)In order to solve the problem that there are many redundant features in the image features obtained by ICFS algorithm,Class Supervised Intelligence Collaboration Feature Selection(CSICFS)is proposed.CSICFS algorithm uses mutual information to calculate the similarity between features and categories,and select more features related to image category information.The features extracted by MELDB are selected to excellent subsets by CSICFS.Finally,this thesis compares the traditional feature description algorithm and the excellent algorithm proposed in recent years on CMU-PIE,Extended-Yale B,Pho Tex and Raw Foot image databases by adding different proportion of salt and pepper noise.The experimental results show that the proposed algorithm can extract and select the optimal feature subset,and improve the image recognition rate in high noise scene.In conclusion,the algorithm proposed in this thesis has some practical application value and academic research value.