Shigueo Nomura

A novel adaptive morphological approach for degraded character image segmentation

This work proposes a novel adaptive approach for character segmentation and feature vector extraction from seriously degraded images. An algorithm based on the histogram automatically detects fragments and merges these fragments before segmenting the fragmented characters. A morphological thickening algorithm automatically locates reference lines for separating the overlapped characters. A morphological thinning algorithm and the segmentation cost calculation automatically determine the baseline for segmenting the connected characters. Basically, our approach can detect fragmented, overlapped, or connected character and adaptively apply for one of three algorithms without manual fine-tuning. Seriously degraded images as license plate images taken from real world are used in the experiments to evaluate the robustness, the flexibility and the effectiveness of our approach. The system approach output data as feature vectors keep useful information more accurately to be used as input data in an automatic pattern recognition system.

Morphological preprocessing method to thresholding degraded word images

This paper presents a novel preprocessing method based on mathematical morphology techniques to improve the subsequent thresholding quality of raw degraded word images. The raw degraded word images contain undesirable shapes called critical shadows on the background that cause noise in binary images. This noise constitutes obstacles to posterior segmentation of characters. Direct application of a thresholding method produces inadequate binary versions of these degraded word images. Our preprocessing method called Shadow Location and Lightening (SL*L) adaptively, accurately and without manual fine-tuning of parameters locates these critical shadows on grayscale degraded images using morphological operations, and lightens them before applying eventual thresholding process. In this way, enhanced binary images without unpredictable and inappropriate noise can be provided to subsequent segmentation of characters. Then, adequate binary characters can be segmented and extracted as input data to optical character recognition (OCR) applications saving computational effort and increasing recognition rate. The proposed method is experimentally tested with a set of several raw degraded images extracted from real photos acquired by unsophisticated imaging systems. A qualitative analysis of experimental results led to conclusions that the thresholding result quality was significantly improved with the proposed preprocessing method. Also, a quantitative evaluation using a testing data of 1194 degraded word images showed the essentiality and effectiveness of the proposed preprocessing method to increase segmentation and recognition rates of their characters. Furthermore, an advantage of the proposed method is that Otsus method as a simple and easily implementable global thresholding technique can be sufficient to reducing computational load.

Improving iris recognition through new target vectors in MLP artificial neural networks

This paper compares the performance of multilayer perceptron (MLP) networks trained with conventional bipolar target vectors (CBVs) and orthogonal bipolar new target vectors (OBVs) for biometric pattern recognition. The experimental analysis consisted of using biometric patterns from CASIA Iris Image Database developed by Chinese Academy of Sciences - Institute of Automation. The experiments were performed in order to obtain the best recognition rates, leading to the comparison of results from both conventional and new target vectors. The experimental results have shown that MLPs trained with OBVs can better recognize the patterns of iris images than MLPs trained with CBVs.

Novel nonspeech tones for conceptualizing spatial information

We propose a novel concept toward interfaces that can provide visually impaired persons with the opportunity to recover the freedom to conceptualize their environment without depending on conventional voice synthesizer systems. Fourteen subjects participated in ten experiments to provide results that evaluated their performances to conceptualize spatial information based on cues in “artificial-sounding” (AS) and “natural-sounding” (NS) tones. The source of AS tones was the digitized sound used by the vOICe Learning Edition, and the source of NS tones was fan noise with analogs in everyday listening. Experimental results revealed that the use of NS tones was essential for improving the conceptualization performance of subjects as the eventual users of novel human–environment interfaces.

Uma nova abordagem utilizando RNA para reconhecimento automático de placas de veículos baseado em pré-processamento de imagens e pós-processamento

This work presents a new system based on artificial neural networks (RNA) techniques to improve the automatic recognition of the number plate from degraded images. The input of the system are images automatically generated by photographic radars installed at the roads in Uberlândia city. The proposed system is composed by a degraded image preprocessing stage and a number plate code recognition stage. New preprocessing methods for image binarization, character segmentation and feature extraction are applied. The RNA model is trained with a set of output patterns that have low similarity between them. This model is used to recognize each character extracted from the number plate. The recognized characters sequence serves as the input data to the postprocessing stage that accepts or rejects that sequence as the result of the number plate recognition. Despite the degraded quality of the input images, experimental results have presented promising results. The output patterns used in the training stage of the RNA model and the proposed postprocessor have significantly increased the recognition rate showing the robustness of the system to process this kind of image.

Orthogonal bipolar vectors as multilayer perceptron targets for biometric pattern recognition

This work proposes the unconventional use of orthogonal bipolar vectors (OBVs) as new targets for multilayer perceptron (MLP) training and test with biometric patterns represented by iris images. Nine different MLP models corresponding to nine different target vectors (including OBVs) have been developed for experimental performance comparison purposes. The experiments consisted of using biometric patterns from CASIA Iris Image Database developed by Chinese Academy of Sciences - Institute of Automation. The experimental results led to conclude that using OBVs as targets for MLP learning can provide better recognition performances rather than using other vectors as targets. Also, the results have shown that MLPs can be trained for OBVs spending smaller number of epochs to achieve relevant recognition rates compared to other types of target vectors. Therefore, the computational load for training MLPs can be reduced and biometric pattern recognition performances can be improved by using OBVs as targets.