Feature Extraction for Neural Network Pattern Recognition for Bloodstain Image Analysis

Feature extraction methods and subsequent neural network performances are explored in this paper. Object recognition method ‘regionprops’ and moment invariants are used to extract basic characteristics from acquired bloodstain images. The extracted features are in return fed into a neural network for the purpose of pattern recognition. The blood drop in the image is first detected using sobel edge detector. After the image has been thresholded and the noise removed, geometric properties of the blood drop is measured with ‘regionprops’. The properties extracted include ‘Area’, ‘Centroid’, ‘MajorAxisLength’, ‘MinorAxisLength’, etc. The seven invariant moments are also extracted from the images. These values are compiled into appropriate input for the neural network pattern recognition function. Two types of neural network modules, cascade forward neural network (CFNN) and function fitting neural network (FFNN) are tested and compared. Several trials have been conducted to determine the performance when using ‘regionprops’ and moment invariants function to extract features. The testing results show that FFNN is better than CFNN at the image recognition in terms of percentage of recognition and resource consumption.