Contour Detection Based On Primaty Visual Model

Edge detection is commonly used in digital image processing, and is a low-level feature extraction means. Although the traditional edge detection algorithms have benefits such as fast, easily to achieve, they did not take into account contextual informations, middle-level informations and high-level informations. Contour detection is one of middle-level feature detection methods in image processing. Compared with the edge detection, contour detection is more robust, and is essential feature extraction methods for target recognition.If the contour detection is regarded as the process of identification, then can we apply visual models to identify the contours in images? The works of this paper are just a series of studies for this idea.In the course of the study, the algorithm of non-classical receptive field inhibition based on contour detection was accessed, which is very similar to the idea of this paper. Compared with the classical edge detection operators, this contour detection algorithm has good treatment effect, especially for the contour detection of objects in the background filled with complex texture. This algorithm using non-classical receptive field inhibition properties can weaken the edge of the background to improve the effect of contour detection. But the accuracy of contour detection of this algorithm is not ideal. There might be two reasons. First, the inhibition for the edges of the background is not strong enough. Second, the algorithm only presented two layer visual model with simple and complex cell layer, but in fact the real structure of the visual cortex is more complex.HMAX model is a kind of visual model with the basic funciton and structure of the visual cortex. The model contains five layers:simple cells layer (S1 layer) and complex cells layer (C1 layer), composite feature layer (S2 layer), complex composite layer (C2 layer) and view-tuned cells layers (VTU layer). But HMAX model is mainly used for target identification. For an input image, after its five-layer treatments, at last, the output is a vector of image features. Therefore, HMAX model is not suitable for contour detection.This paper was aware of both advantages and disadvantages of the algorithm of non-classical receptive field inhibition based on contour detection and HMAX model, using one's advantage to make up the other's disadvantage. Finally the algorithm of this paper was proposed. This algorithm extended the two-layer visual model of contour detection based on inhibition of non-classical receptive field into a five-layer visual model as HMAX model, enhancing the inhibitory effect at the meanwhile. Therefore, this algorithm not only has the basic structure of the visual cortex, but also improves the accuracy of contour detection, particularly suitable for processing images with complex background texture. Based on existing assessment algorithms, through parameter adjustment and partial improvement, the assessment algorithm of this paper was obtained. In this assessment algorithm, evaluation parameters,such as the accuracy of contour detection, the positive error ratio and negative error ratio, were defined. The classical Canny operator and anisotropic inhibition algorithm (one kind of algorithm in contour detection based on non-classical receptive field inhibition) were choosed to be compared with the algorithm of this paper after handling many images in the Grigorescu image library. Through best contour images compared with the other two algorithms', we could see the contour detection images under the algorithm of this paper with less background edges and more integrated contour of the objects. In use of assessment algorithms, assessed parameter table were received. Compared with the other two algorithms, this paper algorithm got a lower mean value of positive error ratio and negative error ratio and a greater accuracy value of P. All three methods processed different types of images with complex background for several times. Then a set of contour detection accuracy values obtained from these methods were assessed by the box and whisker diagrams. The results shows that this paper algorithm has a higher contour detection correct rate.In summary, through detecting the target contour in various types of images with complex background texture, and compared with the Canny operator and anisotropic inhibition algorithm, this paper algorithm improves the accuracy of contour detection, and reduces the detection rate of errors of the background edge. This paper algorithm is a kind of contour detection algorithm with good performance, laying a good foundation for target identification.