Abstract:
Abstract: Image acquisition and reconstruction is one of the key technologies in the development of machine vision technology. With the continuous developments of agricultural automation, image compression and reconstruction of targeted plants play an important role in the detection of fruits and plants. High-speed and high-quality image compression has become a research hotspot. The traditional Nyquist sampling theorem requires that the sampling frequency must be greater than twice of the highest frequency of the signal to completely reconstruct the original signal. The theory of compressed sensing parallelizes the sampling and compression of data, only requires a small amount of signal to accurately reconstruct the original signal, which greatly eases the pressure of storage and transmission. The general compressed sensing methods have problems such as low reconstruction precision and long running time. Aiming to solve these problems, a modified regularized adaptive compressed sampling matching pursuit algorithm based on dog-leg(DLRaCSMP) and compressive sampling matching pursuit (CoSaMP)is proposed in this paper. Regularization method is used in the iterative process to ensure the accuracy of the support set selection. In the iteration process, Dog-Leg least squares algorithm is used to accelerate the convergence speed, and then the residual values obtained in the adjacent reconstruction process are compared. The relative threshold is set to adjust the step size. The reconstruction time is shortened by fast approximation of large step size, and the precision is ensured by accurate approximation of small step size. Kinect 2.0 is used to obtain the color image of the target plant. The brightness, hue and outline features of HSV color space are used to input into the Itti model, and the saliency feature image of the plant is constructed by fusion, which highlights the foreground target and simplifies the complex scene, and reduces the data collection. The test results show that the reconstruction time of the real-time image and salient feature image is about 2.14 and 1.75 s respectively when the sampling rate is 0.50. The average reconstruction efficiency is increased by 75.5% and 77.9% compared with that of the CoSaMP algorithm, respectively. The peak signal-to-noise ratio of the original image and salient feature image reaches 35.16 and 38.93 dB respectively, which improves 6.12 and 5.75 dB compared to that of the CoSaMP algorithm. And the average reconstruction accuracy is increased by 21.6% and 15.5% compared with that of the CoSaMP algorithm respectively.