Abstract AI intelligent detection of colon polyp has been found as a highly popular research direction. Moreover, mainstream research places a focus on how to recognize colon polyp using a better neural network model architecture. The video employed for recognition will be considered the original video output by the endoscope. Through research, it was found that besides the prominent neural network architecture, more excellent video preprocessing algorithms can significantly increase the accuracy of recognition and location for colon polyp. As revealed by the research result, the relative highlight area attributed to uneven illumination and the absolute highlight area attributed to specular reflection are the main factors of the recognition of colon polyp by the neural network. To solve the problem above, all highlight areas are divided into four categories, i.e., the relative highlight area, the large absolute highlight area, the medium absolute highlight area and the small absolute highlight area. This study designs different restoration algorithms in accordance with the nature and characteristics of the respective categories. The relative highlight area can be corrected and restored using the two-dimensional (2d) gamma function. The large absolute highlight area will not be processed since it will not reduce the recognition accuracy of the neural network. The small absolute highlight area has a slight effect on the recognition accuracy of the neural network, so the surrounding color filling method will be adopted to restore the area. The medium absolute highlight area will be restored by the optimized Criminisi algorithm. The test is performed on four neural networks, i.e., the Unet, Unet++, ResUnet and ResUnet++. After the sample is processed by this algorithm, the results show that the recognition accuracy of colon polyps by four kinds of neural network is significantly improved. Compared with other image restoration algorithms that take tens of seconds, the image restoration algorithm in this study takes less than 90 ms, which obviously reduces the time, and can basically meet the real-time requirements of AI intelligent detection.
展开▼
