Gastric cancer is one of the most common digestive tract malignant tumors in the world. Japan, South Korea and China are high incidence of gastric cancer in Asia. There are about 400 thousand new cases in China each year, accounting for 42% of the total number of cases in the world[1]. With the development of artificial intelligence, especially deep learning, people pay more and more attention to computer-aided diagnosis, where some progress has been made in the study of gastric cancer slice images. By extracting and classifying the features of the gastric slices, computer-aided diagnosis system can judge the normal and abnormal of the gastric and help doctors to make a diagnosis. In essence, the above process is to divide gastric slice images into two categories, cancer and non-cancer. The classification results have a great relationship with features extraction of images and performance of classifier. In previous work, there are two kinds of manual bottom feature extraction method: one is interest points detection, the other is dense extraction [3]. Interest points detection algorithm selects obvious feature pixels, edges, corner points or blocks by some criterion, which generally has the geometric invariance and small computation overhead, such as the Harris corner detection, Features from Accelerated Segment Test (FAST), Laplacian of Gaussian (LoG), etc. Dense extraction method extracts a large number of local features from the fixed step length and scale. Although a l a r g e n u m b e r o f l o c a l f e a t u r e s h a v e higher redundancy, the feature information is more abundant. This method will achieve better result compared with the feature extraction method based on interest points. The common local features include Scale-Invariant Feature Transform (SIFT), Histogram of Oriented Gradient (HOG), Local Binary (LBP), etc. However, a popular view in recent years is that using the low level feature descriptor as the first step of visual information processing often loses useful information too early. Directly learning feature descriptions related to task from image pixels is more effective than manual features.