Design And Implementation Of Real-time Detection System For Specific Small Targets

Target detection has been one of the important technologies in the field of image processing since the last century.It has been widely used in military target strike,intelligent video surveillance,and aerospace infrared warning system.In one of these applications,the region of interest exists as the form of small target,which only occupies a small image area,lacks texture,and has blurred edges.Only fewer features can be extracted.At the same time,due to different application scenarios and different environmental complexity,small targets usually exhibit different characteristics.Therefore,the detection of small targets from complex background has become a hot spot and research puzzles for the relevant researchers.At present,research on small targets often requires specific processing for specific scenarios to achieve better detection effect.It is even more difficult to achieve real-time requirements while ensuring the detection effect.In such a situation,this thesis considers the detection problem of a small target with periodic variation characteristics in the time domain from the actual scene,and studies the design and implementation of the real-time detection system for this specific small target.In order to balance the accuracy and real-time performance of the detection system,this thesis splits the small target detection process into two parts: sub-block level target coarse positioning and pixel level target fine positioning.Firstly,through the coarse positioning part,the approximate area where the target may exist is quickly searched,and then the target fine positioning part is used to accurately detect the real target position.In the coarse positioning part,based on the interframe difference method that is simple but has poor detection performance,this thesis proposes a new sub-block level feature difference method.In the fine positioning part,based on the local contrast measure(LCM)that is complex but has better detection performance,thesis proposes fast multiscale local contrast measure(FM-LCM).The main work and innovations of this thesis include the following three points:In the sub-block level target coarse positioning part,this thesis proposes a new sub-block level feature difference method based on the inter-frame difference method and the time domain characteristics of a specific small target.The method requires sub-block division and feature extraction for the image to be detected.On this basis,the feature difference is used to screen out the suspicious target block.On the one hand,image sub-blocks are used instead of pixel points for difference,which can reduce the sensitivity of the algorithm to background edges and noise,and also convert the brightness characteristics of small targets from pixel level to sub-block level.On the other hand,sub-block partitioning can screen out a number of suspicious target blocks with regular shapes,which is convenient as an input for subsequent detection links.In the pixel-level target fine positioning part,this thesis analyzes the advantages and disadvantages of the LCM small target detection method,and proposes a new FM-LCM method to further eliminate noise interference and achieve the purpose of accurately detecting small targets.The FM-LCM method proposed in this thesis optimizes the filtering window structure of LCM,improves the response processing of the algorithm,and solves the problem that LCM can not adapt to multi-scale target detection,and reduces the computational complexity while ensuring the detection accuracy.Finally,this thesis constructs the algorithm framework and complete process of the target real-time detection system,and completes the code migration and optimization work on the hardware.The software and hardware joint debugging is carried out to verify the real-time performance of the system.