Plant Seed Identification

Plant seed identification is routinely performed for seed certification in seed trade, phytosanitary certification for the import and export of agricultural commodities, and regulatory monitoring, surveillance, and enforcement. Current identification is performed manually by seed analysts with limited aiding tools. Extensive expertise and time is required, especially for small, morphologically similar seeds. Computers are, however, especially good at recognizing subtle differences that humans find difficult to perceive. In this thesis, a 2D, image-based computer-assisted approach is proposed. The size of plant seeds is extremely small compared with daily objects. The microscopic images of plant seeds are usually degraded by defocus blur due to the high magnification of the imaging equipment. It is necessary and beneficial to differentiate the in-focus and blurred regions given that only sharp regions carry distinctive information usually for identification. If the object of interest, the plant seed in this case, is in- focus under a single image frame, the amount of defocus blur can be employed as a cue to separate the object and the cluttered background. If the defocus blur is too strong to obscure the object itself, sharp regions of multiple image frames acquired at different focal distance can be merged together to make an all-in-focus image. This thesis describes a novel non-reference sharpness metric which exploits the distribution difference of uniform LBP patterns in blurred and non-blurred image regions. It runs in realtime on a single core cpu and responses much better on low contrast sharp regions than the competitor metrics. Its benefits are shown both in defocus segmentation and focal stacking. With the obtained all-in-focus seed image, a scale-wise pooling method is proposed to construct its feature representation. Since the imaging settings in lab testing are well constrained, the seed objects in the acquired image can be assumed to have measureable scale and controllable scale variance. The proposed method utilizes real pixel scale information and allows for accurate comparison of seeds across scales. By cross-validation on our high quality seed image dataset, better identification rate (95%) was achieved compared with pre- trained convolutional-neural-network-based models (93.6%). It offers an alternative method for image based identification with all-in-focus object images of limited scale variance. The very first digital seed identification tool of its kind was built and deployed for test in the seed laboratory of Canadian food inspection agency (CFIA). The proposed focal stacking algorithm was employed to create all-in-focus images, whereas scale-wise pooling feature representation was used as the image signature. Throughput, workload, and identification rate were evaluated and seed analysts reported significantly lower mental demand (p = 0.00245) when using the provided tool compared with manual identification. Although the identification rate in practical test is only around 50%, I have demonstrated common mistakes that have been made in the imaging process and possible ways to deploy the tool to improve the recognition rate