Research On Non-uniformity Correction Technology Of Infrared Focal Plane Arra

As a visualization technology with strong anti-interference ability,infrared imaging technology is widely applied in military confrontation,night driving,disease diagnosis,forest fire,weather prediction,and anomaly detection.However,due to the limitation of the processing technology,the infrared image has the characteristics of strong nonuniformity,low contrast,and serious blind element,which are not conducive to visual observation.To improve the imaging quality of infrared images,this thesis mainly researches the characteristics of the nonuniformity,blind pixels,and the contrast of the infrared images.The main contents of the thesis include:(1)The basic principles and implementation steps of several non-uniformity correction algorithms are introduced.Then the advantages and disadvantages of these algorithms and optimization ideas are elaborated through algorithm principles or experimental simulation.(2)The correction algorithm based on registration is studied.The algorithm was optimized from the selection of registration algorithm,improvement of registration accuracy,local space,and suppression of ghost phenomenon.The experimental results show that when the field of view moves rapidly,the proposed method can remove the striped nonuniformity noise without producing the visible ghost phenomenon.Compared with the traditional method,the roughness index of this method is reduced by about 1.4%.The non-uniformity correction algorithm based on the random number is proposed to remove the spatial low frequency and high-frequency non-uniformity noise in the scene.The method uses random numbers to construct the expectation function,so the algorithm complexity is very low and the real-time performance is very high.Experimental results show that compared with other methods,this method can better remove spatial low-frequency nonuniformity noise such as "pot cover" noise.Compared with the traditional neural network method and the registration method,the roughness index of this method is reduced by about 31% and 35% respectively,reflecting better noise removal ability.It is worth noting that although this method can remove both low and high-frequency nonuniform noise in the airspace at the same time,the ghost will be introduced when the scene is still or moving slowly.(3)The correction algorithm based on a neural network is studied.The judgment mechanism of algorithm convergence is introduced and a smoothing function is designed to eliminate ghosting.Experimental results show that the proposed method has low resource consumption and high real-time performance.The running time of the algorithm for testing the324×256 infrared image dataset on Intel COOL I5 8256U-CPU is not less than 100 frames/SEC.The structure similarity index of this method is 24% higher than that of the traditional neural network method,and it can keep the details of the image well while removing the noise of inhomogeneity.In addition,compared with other non-uniformity correction algorithms based on scene time-domain iteration,this algorithm can achieve non-uniformity correction under the static and slow motion of the field of view without introducing ghosts.An image detail enhancement algorithm based on guided filtering is studied.A neural network method based on guided filtering and a one-point correction method based on time-domain iteration was designed to remove the striped nonuniformity noise and the "pot cover" noise in the image respectively.Guided filtering is used to separate the fundamental frequency information and detail information of the image,and an improved contrast constrained histogram equalization algorithm is designed to map the 14 bit fundamental frequency subgraph to the 8bit image space.The experimental results show that the proposed method can effectively remove the striped nonuniformity noise and "pot cover" nonuniformity noise,and enhance the detail features of the image.In terms of visual effect,entropy enhancement measurement and structural characteristics,this method is superior to other classical image detail enhancement algorithms.(4)The nonuniformity correction strategy based on deep learning is studied.The U-Net network is trained to achieve non-uniformity correction,and the network structure is optimized to improve the detail retention ability of the model.The design scheme of the lightweight network is studied and the convolutional pruning strategy is used to reduce the complexity of the model.Finally,the trained model is transplanted into the Cambrian MLU development board platform.The simulation results show that the proposed algorithm is 5% and 4% higher than the traditional U-NET algorithm in terms of peak signal-to-noise ratio and the structural similarity index,respectively,and can better protect the detail features of the image.After compression pruning,resource consumption is reduced by 30.6%,peak signal-to-noise ratio,and structural similarity index are reduced by 7% and 6%,respectively,compared with before pruning.The compression pruning strategy only loses a small amount of network accuracy and greatly reduces the model overhead.