Fingerprint pre-processing and feature engineering to enhance agricultural products categorization

Highlights

• We investigate the fingerprint feature extraction.

• And optimize the performance of fingerprint login authentication system.

• Introducing a new feature extractor based on binary image through a brief thinning process.

Abstract

Among many biometric technologies, fingerprint identification is the earliest, most pervasively applied, and cheapest biometric technology. It seamlessly corporates integrates sensors, biotechnology, electronic technology, digital image processing and pattern recognition. Fingerprint based identity authentication has expanded from the traditional criminal investigation to e-commerce, attendance, access control, credit card, endowment insurance system and other fields, and has become the popular biometric technology. As for the existing matching algorithm such as feature-point-based or image-based and so on, the accuracy and speed of fingerprint automatic identification are far from the satisfaction of many practical applications. In this work, we investigate the fingerprint feature extraction, and optimize the performance of fingerprint login authentication system by introducing a new feature extractor based on binary image through a brief thinning process. Aiming at the problems of low flexibility, high cost and low efficiency of algorithm in embedded chip of current fingerprint identification system, the automatic fingerprint identification system is comprehensively analyzed. Our proposed method can improve the design and optimization of the automatic fingerprint identification system. Our method can improve the accuracy of fingerprint image recognition in the automatic fingerprint identification system. Meanwhile, it can also improve the adaptability and stability of the automatic fingerprint identification system in various domains. Also, our method can substantially improve the feature extraction of agriculture products. Thereby, the classification task can be enhanced accordingly.

Introduction

In fingerprint recognition, as the foundation of fingerprint matching, the preprocessing of fingerprint image and feature extraction of fingerprint image play a significant role in the entire fingerprint recognition process. Fingerprint image segmentation, fingerprint image enhancement, fingerprint image binarization, fingerprint image refinement, etc. the binary fingerprint image is to transform gray fingerprint image into binary image represented by only 0 (representing ridge sub graph) and 255 (background sub graph). This operation transforms each gray image into black and white image. Ridge thinning processing of fingerprint image continuously erases the edge pixels of the image without changing the topological connection relationship of the image pixels. This operation converts the fingerprint image with uneven ridge thickness into the central line image of single pixel wide stripe as much as possible. The ridge thinning process of fingerprint image not only increases the processing time, but also introduces more pseudo features during the thinning process. There are many burrs, short lines, breakpoints and holes in the original skeleton in the fingerprint image after the binary image is refined. Further noise removal is required. In order to simplify the algorithm, the binary image refinement is not conducted in such design. In the subsequent feature extraction, the binary image based feature extraction algorithm is adopted. This not only avoids the pseudo feature points in the refinement process, but also shortens the entire processing time. This will not affect the effect of feature extraction. In the preprocessing stage of fingerprint image, the effective fingerprint region segmentation and direction field calculation are combined with the direction field calculation. Using the nonlinear diffusion model image enhancement algorithm, the fingerprint filtering spreads along the ridge line, that is, the feature extraction of the broken fingerprint ridge is a primary operation in image processing. That is to say, the first level of operation processing of an image. It checks each pixel to determine whether the pixel represents a visual feature. The feature extraction of an image is the fundamental and key feature for image modeling, especially for fingerprint images. According to the FBI record, human fingerprint is divided into eight types. However, because of these eight types, the fingerprint image is divided into 8 types. The difference of some fingerprints is very small and thus it is difficult to distinguish them accurately by a computer. The National Bureau of standards and Technology (NBST) issued the standard fingerprint database. This database has becoming the standard of many automatic fingerprint classification algorithms. This kind of fingerprint feature is popularly used in one-to-many matching of large fingerprint database, that is, fingerprint classification system. The classification based on the entire fingerprint feature can be leveraged to distinguish different fingerprint to some extent. But these information is insufficient to distinguish all different fingerprints. In this way, they are typically used in classification and retrieval. The local features of fingerprint refer to the nodes on the fingerprint. Two fingerprints typically have the same overall characteristics, but their local features, nodes, cannot be identical. Fingerprint patterns are usually uncontinuous and straight. Instead, there are frequent interruptions, bifurcations or discounts. These breakpoints, bifurcation points and turning points are called “nodes”.

This is the characteristics of these nodes encoding the identification information of fingerprint uniqueness. The local features of fingerprint denote the fingerprint nodes. The error recognition rate of the simple feature matching algorithm is high, and the feature of detail points indicates that lots of ridge information is lost. These information is also very informative for distinguishing features among different fingerprint, and thus limits the performance of the fine node method. The texture feature of fingerprint globally describes the surface properties of the scene corresponding to the image or image area. However, since the texture is only a feature of the object surface, it cannot comprehensively reflect the essential attributes of the object. Therefore, only using texture features cannot obtain high-level image content. Texture-based feature description method can retain rich ridge information. To our knowledge, it can overcome the difficulty of difficult extraction of the details of poor quality area. Noticeably, this method needs a reliable reference point, which is difficult to deal with the inelastic deformation of fingerprint. In order to improve the system recognition accuracy, the two algorithms are combined with the complementarity of the two algorithms. A backward–forward scanning feature extraction algorithm based on the comprehensive feature (detail point is part of texture feature) is formulated. This feature extraction can be applied to agriculture leaves categorization task.

The collected fingerprint image will be changed due to the interference of many external factors. In order to enhance the reliability of fingerprint extraction, image enhancement is required before extraction (as shown in Fig. 2). The collected gray level of fingerprints will be enhanced by filtering, the initial binary image is obtained, the image is refined, and a new binary image is finally obtained. Based on the two images, the following operations are made. The relevant ridge structure is conducted by simple rules, namely the bridge between ridges, the obvious short line, and the short ridge line burr. The local ridge line is fitted by leveraging the primary curve and secondary curve. Subsequently, the fitting curve is calculated from the end of ridge line. Different rules are required to deal with different situations. The latest binary image is obtained by the above processing. The ridge points and fine nodes are extracted. The effect of image enhancement significantly influence the performance of fingerprint recognition system. After the enhancement, the fingerprint feature extraction step is conducted, wherein the gray fingerprint image is transformed into black and white image. The direction array is obtained as the basis of using classification system. Generally, when processing the two fingerprint images, the difference between the two images is determined using the point patterns of the two images. The difference of fingerprint features is from two aspects: single point and detail feature. The single point includes two types of center point and triangle point. The detail points are divided into bifurcation point and ridge end point in fingerprint image topology. The distance between the center point and the triangle point is relatively fixed, and will not change with the rotation, scaling and transformation of fingerprint image [1].

In summary, the key technique of matching fingerprint is to compare pairwise fingerprint images, and subsequently determine whether the two fingerprints belong to the same person using the similarity of their features. When matching pairwise point patterns, we formulate the detail point sets from one of fingerprint images beforehand, and subsequently compare them with the points set of the images. During this process, the similarity between fingerprints is calculated by leveraging the endpoint and the separation point. If the point mode of two images is compared through rotation, translation, scaling and other transformation, the similarity between two fingerprint images match can be calculated accordingly. Fingerprint image processing algorithm needs the image direction filtering. Based on the ridge line direction, the direction of block is obtained, based on which the point direction map is divided to get N $\times$ N blocks. We calculate the distribution probability of points in the block, and obtain the direction of the maximum distribution probability of points in the block. This is the direction of the block. Through these operations, we can acquire the direction diagram that is handled by leveraging the filter in the horizontal direction. Herein, the filter in other directions is calculated by the filter in the horizontal direction through rotation. The weight functions involved in the directional filter are two groups, corresponding to the separation filter and the average filter. In the process of block pattern denoising, such direction filter is utilized to calculate the block direction [2].