Image Segmentation And Recognition In Seed Classification And Inspection

Seed inspection is a process that analyzes and evaluates seeds using scientific method to identify quality of the seeds. The quality of seeds could be generally divided into two aspects: one is variety quality, which indicates the inner value of seeds, such as variety purity and seed authenticity; the other is sowing quality, which influences the field harvest and usually includes the seed purity, plumpness, percentage of germination, vitality, moisture content, specific weight, volume weight, infection rate, etc. According to the two aspects of seed quality, the seed inspection could also be generalized as: variety classification, which mainly revolves seed purity inspection, and quality inspection, which emphasizes the robustness and healthiness inspection. The manual seed inspection is both time-consuming and tedious, and the seed inspector suffers from weariness and continually decreasing accuracy.Computer Vision is an intelligent technology emerged in the last few decades and has been extensively applied into production, research and our daily life. It is especially demanded in the circumstance which is repeated, monotonous and depending on vision, like quality inspection and evaluation in batch production. The machine could work with higher speed and accuracy than human inspectors. Seed classification machine based on computer vision usually includes software component and hardware component. The software is composed of image processing, analysis algorithm, seed pattern recognition, etc. The hard ware component includes seed deliverer, image acquisition and computer system etc. In this paper, we mainly discussed the software part of the seed analyzer, and presented a variety classification and inspection method.In the variety classification, seed images are processed and analyzed and the seed geometrical and physical features are extracted. The feature array is the mathematical model, and also called pattern. To choose appropriate features for the seed model building, could bring compacted seed cluster distanced from each other. Using experimental method, 24 parameters are chosen, which are area, length, mean width, max width, center of mass, max width position, rectangularity, compactness, light information, texture, edge and Fourier. Under model building mode, the software processes a batch of seed and saves feature vector for each seed observation, and finally builds standard variety classification model. Under working mode, each new seed observation is compared with standard model and is classified into the nearest variety.In the seed quality inspection, the software evaluates the seed quality in five aspects: volume, plumpness, healthiness, broken rate and impurity rate. Used several features to quantify those aspects on the basis of expert experience: area in 2D image stands for the seed volume; rectangularity could represent seed plumpness; when seed is infected by virus or insects, the color is different from healthy seeds; broken seeds have unsmooth surface that results in more edge area and rougher texture; seed impurity rate is the division of good seed and total seed, and could indicate the farmable fraction of total seed. Therefore, seed quality could be quantified directly using features of seed model, which are area, rectangularity, RGB light intensity, edge area and texture descriptors. Seed impurity rate could be derived from seed sample statistical information. After determination of quality features, built the membership function for fuzzy decision. When a set of seeds are processed, the membership is computed and quality inspection result is derived.The software presented in this thesis could recognize five different seeds, namely wheat, barley, rye, oats and corn. 1000 observation of each variety were tested and the overall classification accuracy rate is above 92%, and the quality inspection could correctly describe the quality condition of the 5 sets of seed. The software is user-friendly, the whole process is monitored through the human computer interface, and all critical parameters could be modified directly. The software is adaptive as new seed model could be built at any time. This figured the geographic and time limitation of seed model as seed condition is greatly influenced by different environment conditions. The software also has good expansibility as the features introduced could describe most known seed classes.In sum, this thesis presented a seed classification method using image segmentation, built seed model using representative features, took account imaging angle issue and realized more accurate mathematical description by further classification of seed model. The seed quality inspection method could more comprehensively describe quality condition, is especially applicable to agricultural industry. The software is stable, reliable, accurate and effective, could replace human power in the seed classification and inspection, and has practical value for industries that use seed as raw material.