Research Of Target Detection Based On Wavelet Transform

Target detecting technique has extensive applications in numerous fields, andthe research on this technique has important theory and practice value. In thisPaper, systemic analyses on target detecting technique and its latest progress indomestic and international is carried out, along with underway research project.On the bases of classic target detecting algorithms, new target detectingalgorithms based on wavelet transform is researched, especially on targetdetecting algorithm of lift technique and the hardware realization of targetdetecting algorithm. The principle research work and achievements of the paperare as follows:1. According the characteristics and requirements of target detecting,analyzed the corresponding wavelet transform theories. After introducing thecontinuous wavelet transform and multi-resolution analyses and MALLAT fastdecompose algorithm, which is important in wavelet transform, the paperdescribed the 2-D multi-resolution analysis composition and decompositionprocedures of image. Especially in the advantage of structure and characteristicof wavelet transform based on lifting method and the integer wavelet transformwith lifting method. The theoretic basis of target detecting algorithm is settled.2. Introduced the algorithm of fast multi-scale self-adapted thresholddetecting object fine edge. At the same time, analyzed the advantage of wavelettransform in noisy remove and image enhancement etc in image preprocess.Combined the advantages of multi-resolution analyses of wavelet transform inedge detect, noise remove, and image enhancement, to detect the edge character oftarget exactly.The wavelet transform can analysis the oddity position and intension ofsignal. The paper analyses the feasibility and advantage of wavelet transform onedge detecting, and detects target through tracing the mode maximum curve ofwavelet transform in fine scale. In the paper, the edge detecting method ofapplying wavelet mode maximum is analyzed, especially the method of multiscale edge detecting. Several edge detecting methods of applying wavelettransform local mode maximum are compared. Meanwhile, by applying thewavelet analyses, how to remove noise and enhance image in edge detecting,enhance the useful information, and detect the fine edge of target in image, locatethe target exactly is discussed. The emulated experiment results show that theedge detecting algorithm of multi-scale self-adapt threshold by applying B samplewavelet can detect the target fine edge very well, especially to the special targetimage severely polluted by noise.3. On the research bases of the ways of applying wavelet transform to detectsmall weak target, in order to improve the efficiency of detecting small weaktarget, presented the multi-scale wavelet transform and energy chiasma algorithm.The weak small target is the case where target is only a few pixels and haslow signal and noise rate in image plate. According to the different characters, theweak target can be divided into two types, one is low contrast target, e.g. grayscale weak target, the other is target that few pixels, e.g. energy weak target (smalltarget). Gray scale weak target is present by the SNR of target image. On theprinciple of wavelet transform, by applying different scale of wavelet transform toa image, a series of similar image would be generated. The algorithms to use thischaracter are SS(Sequential-Scaling) algorithm, which is similar to energy detectalgorithm. The principle is based on the fact that index is bigger after wavelettransform on target, compute the energy of index and compared with threshold tofind target. The MCWT (Multi-dimensional Continuous Wavelet Transform)algorithm according the fact that the scale and frequency of wavelet hascorresponding relations, and find the target with the peak value of scale anddirection angle. The advantage is it can decline the false alarm rate effectively, butit is very sensitive to target position in image. The wavelet multi-scalecorresponding distance image detect algorithm depends on the fact that target isonly a small part in whole image. Generating character vectors by using everywavelet indices, then detect the target according the different vector size of targetand background. The main defect is its bigger computing amount. The waveletclass Fisher's algorithm detects the target using the wavelet character generatorand classifier, but the algorithm is too complex and has enormous computingamount and is hard to real application. By wavelet decomposition the target signaland noise are spread to different frequency belt. Because the noise in differentfrequency belt is not corresponded, the energy chiasma method can restrainoverall noise. But the signal has bigger energy than noise in every frequency belt,e.g. corresponded. The result of energy chiasma makes the energy value of edgeincreased and energy of noise weakened. With this, the purpose of restrainingnoise and enhancing target can be achieved. The experiment results show that thecombined algorithm of multi-scale correspond and energy chiasma iscorrespondingly simply and can detect weak small target whose contrast is lessthan 3% successfully.4. The use of lifting wavelet transform in special target detection is discussed.The reason to present this algorithm is IWT provides a speed realization.Compared with the classic Mallat algorithm, the amount of computation isdecreased to a half. The algorithm realizes the original position computation, andno excess memory is needed in the computation process. The positive and inversetransform is reverse order relation, and is easy to perform integer wavelettransform and to deal with edge problem. The lifting wavelet transformabandoned the condition of 2-D movement and flex but regained wavelet has allthe advantages of first generation wavelet. From the principle of wavelettransform, multi-scale wavelet transform can be realized by lifting waveletcompletely. The edge detect problem we studied in the beginning of the paper isjust realized by the essence of lifting wavelet. Besides, by the Sweldens principle,the lift item can be seen as an adjustable argument. The target detect algorithm bylifting wavelet we presented utilize the character of lift item. By learning andtraining, let it has the character of target and use the character to construct newlifted wavelet filter. Through Compute the character point containing target inreference image, e.g. the obtained position which has the bigger high frequencypart computed by the initial filter, to realize the target detect. At the same time,detecting target with the integer wavelet transform by lifting method is discussed.The emulated experiment results show that the lifting wavelet algorithm candetect out target very well. The proposed target detect algorithm based on liftingwavelet is hopeful be realized in practical real-time tracing systems.5. Applied several wavelet algorithms on dual-DSP hardware platform. Useassemble language and pipeline technique of DSP to optimize the programs. Carryforward the advantage of parallel structure in DSP and let the complex waveletalgorithm can operate real-timely. In the DSP system, the wavelet transform ofMallat(FBS) algorithm, the wavelet transform(LS) based on lift method, andinteger wavelet transform(IWT) based on lift method are individually realized,and analyzed the operation speed and experiment results of the algorithms fromthe system structure of DSP, which establishes the experiment foundation for thehardware realization of the algorithms. The hardware realization results of wavelettransform provides the experiment data and analysis basis for the contents in thepaper applying on real system.