Feature extraction of shrimp with computer vision.

摘要

Features were identified and evaluated for a machine vision based shrimp deheader. The two major steps in the extraction process were (1) feature identification and (2) feature retrieval. Spectral and morphological features were examined for fast head identification of the shrimp, Penaeus vannamei. Spectral features included transmittance, reflectance, and emittance of the shrimp. Reflectance and transmittance were evaluated between 400 and 1000 nm. Morphological features included five morphometric variable measurements provided by biologists and a new feature that was generated by a thinning operation on the shrimp images. The features were retrieved with machine vision techniques. Major task in spectral feature retrieval were the design and evaluation of (1) adaptive thresholding techniques, (2) classifiers, and (3) scanning algorithms to locate the cutting reference point for the deheading process.;The intensity averaging thresholding technique was developed to determine threshold values adaptively for color varying shrimp. This technique was effective and fast in segmenting shrimp images. The recognition operation was accomplished with a two step procedure: (1) identification of the hepato pancreas (HP) and (2) location of the cutting reference point. Two simple classifiers were evaluated to classify the HP blob from video noise in the shrimp images. A scanning algorithm based on the medial axis transform (MAT) concept was developed for faster labeling of the cutting reference point.;It was difficult to measure biological morphometric variables with the machine vision system. A new morphological feature, the ratio between the spine length and the cut length, was generated and retrieved. The retrieval of this feature was accomplished by obtaining the skeleton of the region by a thinning algorithm.;The best feature used for shrimp head identification was the difference of the overall transmittance between the head and the tail. The standard deviation of predictions made with the spectral feature ranged from 2.766 to 4.600 mm with an identification rate of three shrimp per second.