Research On Key Technologies Of Vehicle License Plate Recognition System On Complex Conditions

In recent years, the Vehicle License Plate Recognition (VLPR) System has been widelyused in highway toll management and over speed monitoring, urban intersection traffic vio-lation monitoring, car park management, residential vehicle management, and is a significantcomponent of the Intelligent Transportation Systems (ITS). However, an advance weathercondition, such as fog or haze, can usually result in low contracts of the images captured byimaging devices. Moreover, an inaccurate lens focusing, or the high speed of vehicles, mayalso cause the images blurred. Since the image degradations, the identification accuracy ofthe VLPR System decreases, which not only brings many inconveniences to the traffic man-agement, but also poses potential threats to the public safety. Therefore, it is necessary tocarry out studies of VLPR System on this kinds of complex conditions.In this paper, the degraded vehicle image was classified in three categories, namely, fogimages, out-of-focus blurred images and motion blurred images, and studies of restoring thisthree kinds of degraded images were carried on:For images degraded by fog, we analyzed the atmospheric scattering model and foundthe contrast characteristic of degraded images. Basing on this characteristic, we proposed amethod to distinguish whether the image has been degraded by fog and classify the fog indifferent degrees. Considering that the objects in scene locate at different depths of field,leading to different degradation degrees, we presented an adaptive method to select thethresholds of contrast limited adaptive histogram equalization (CLAHE) based on atmos-pheric transmissivity estimation. Finally, the performances of three most common used sin-gle image haze removal algorithms were compared. In order to take full advantages of thesefog removal method, we designed an adaptive vehicle image defog process which couldselect diffident algorithms for images of different degradation degree.For out-of-focus blurred images and motion blurred images, the first step is to detectwhether and where the input image has been blurred. We introduced power spectrum slope and gradient histogram span, and proposed a method to detect blur regions of vehicle imagesbased on the two measures. Then, we to discriminate the types of blurs inside those detectedblur regions according to the shapes of their cepstrums. For an out-of-focus blurred imageregion, we calculated the autocorrelation of the derivation image to identify the out-of-focusradius, and for a motion blurred image region, we adopted cepstrum analysis way to identifyits motion direction and motion length. Finally, in aim to improve the accuracy of the motionblur PSF estimation algorithms for noisy vehicle image, we presented a pre-processingmethod to enhance the straight dark lines structure of the Fourier spectrum based on waveletreconstruction and ridgelet transform. The experimental results demonstrate that the spec-trum pre-processing method proposed enhances the straight lines structure satisfactory forboth noise free and noisy motion blurred images, and improves the accuracy of the subse-quent PSF identification algorithms based on cepstrum.In addition, we completed the works to design and implement the VLPR System calledqVLPR, which is adapted to some complex conditions. This system has the ability to deter-mine if the input vehicle image is degraded with identifying the degradation type (fogged,out-of-focus blur or motion blur) automatically. Then, it adopts the corresponding imagerestoration algorithm to improve the image quality and make the image region(s) of licenseplate clear. The pre-processing step improves the accuracy of the system’s VLPR moduleunder complex conditions.