Firstly, the history and actuality of machine vision inspection system was introduced in this paper, then the common defects of sausage packages was introduced in detail. Finally, the main content of this article was introduced including 3 parts: the first part is image stitching of the sausages pictures based on binocular vision, the second part is the detection of the buckle-losing defects of sausages, the third part is the classification of different brands of sausages. Obtaining high resolution images of sausages is the premise of detection of sausages package defects. In order to improve the image resolution, reducing the distance between the camera installation position and the conveyor belt will make the camera unable to capture the complete images of sausages. In order to solve this problem, the principle of binocular vision measurement system and the theory of image stitching algorithm was analyzed in detail. Using binocular cameras to take the left and right parts of the sausage and stitch them into a complete sausage image. buckle-losing defects are common defects of sausages package. For detecting this kind of defects,the idea of using shape feature to identify this kind of defects was proposed in this article. After thorough study of different shape feature descriptors and the characteristics of buckle-losing defects of sausages, Fourier descriptors were adopted in this article. Then use Euclidean distance method and SVM (Support Vector Machine) to identify buckle-losing defects respectively and compare the recognition accuracy of the two methods. Compared with the existed recognition methods, the recognition method based on Fourier descriptors and SVM greatly improves the recognition accuracy rate of buckle-losing defects of sausages. In the process of sausages production, due to machine error and workers' negligence, one brand of sausages will be mixed into another brand of sausages. The color feature is the most prominent feature for the identification of the wrong brand of sausages. In this article, the difference of sausages with different brands were well studied, and the H (chroma) component of HSV color space was extracted from sausages pictures. Based on H component, the mean, fuzzy entropy, entropy and anisotropy were calculated and combined to a four-dimensional feature vector. The multi-layer perceptron is trained with the feature vectors calculated above, and then different brands of sausages were identified using the trained model.