With the development of computer image process technology, traditional manual selecting has gradually been replaced by automatic selecting equipment in the selecting of the colonies. The core of the automatic screening instrument is the colony image recognition module. The key to the colony image recognition lies in the image segmentation technology. This paper proposes an improved watershed segmentation algorithm. The algorithm first uses Gaussian filtering to remove image noise, then performs morphological processing on the denoised image, then performs chamfer distance transformation to obtain the colony distance image, and then fills the cavity information by morphological method, and then performs the labeled region. The watershed is segmented, and finally the region is merged to gather the similar regions, so that the final segmentation image is obtained. The accuracy of colony image segmentation using the improved watershed algorithm proposed in this paper is 93.4%, while the traditional watershed algorithm has a segmentation accuracy of 75%, it can be seen from the comparison with the traditional watershed that the improved algorithm better suppresses the over-segmentation of the traditional watershed and greatly improves the accuracy of recognition.