International Journal of Applied Pattern Recognition (4 papers in press)
A Simple and Practical Review of Over-fitting in Neural Network Learning
by Oyebade Oyedotun, Ebenezer Olaniyi, Adnan Khashman
Abstract: Machine learning inspired by artificial intelligence describes systems that can adapt, update and record adaptation of internal parameters during a phase known as training. The current states of these internal parameters describe how much experiential knowledge has been acquired by such systems. The minimum required knowledge by such systems in order to be considered applicable is estimated using a cost function which relates the deviation of the actual internal parameters states from the optimal states for sufficient learning of a specific task; the cost function is generally optimized during training. The intelligence of such trained systems can be considered as its capability to perform well upon using unseen data drawn from the same distribution as with training examples; a phenomenon known as generalization. In many situations, of course, it is desirable that a trained system achieves zero classification error on the training examples while tuning its internal parameters for a task, but the amount of generalization power that is lost while enforcing such a learning constraint on the system is quite important. This paper limits its scope to reviewing the consequences of enforcing such learning constraints, and to address the question of where to draw the line on whether a machine has learned a smooth mapping function or largely memorized training data from a practical perspective. For our experiments, we consider handwritten character and digit recognition applications using a publicly available datasets.
Keywords: Neural network; memorization; neural networks; over-fitting; generalization.
Object Recognition with Discrete Orthogonal Hahn Moments
by Fatima Akhmedova, Simon Liao
Abstract: In this research, we have analysed the image feature descriptive properties of discrete orthogonal Hahn moments, and proposed a new object recognition scheme with three modes of Hahn moment descriptors, global, local, and hybrid, respectively. For each mode, we have employed the four highest variances values from ten lowest order of Hahn moments to compose a four dimensional feature vector.To clarify our new scheme of using discrete orthogonal Hahn moment characteristics, we utilized a set of 6; 763 Chinese characters defined in Chinas national standard GB2312, with the font of song, as the testing object set. Each of the three Hahn moment modes has performed very well, while the experimental results of utilizing the three Hahn moment modes are independent from each other.
Keywords: Discrete orthogonal Hahn moments; Object recognition; Chinese character recognition.
Design of Novel Post-Processing Algorithms for Handwritten Arabic Numerals Classification
by Ayatullah Faruk Mollah
Abstract: The process of handwritten character recognition is more challenging in comparison with printed character recognition because of individual specific differences in writing style. Although, extensive research has been carried out, classification accuracy is yet to be improved for practical applications. In the current work, some novel post-processing techniques have been presented for improving recognition performance of isolated handwritten Arabic numerals using multi-class support vector machines. The system is trained and tested using a database of 3000 handwritten Arabic numeral samples. A set of 60 features consisting of 24 shadow, 16 octant-centroid and 20 longest-run features is employed in the current work. Post-processing improves the classification accuracy from 98.10% to 99.30%. The samples that still remain misclassified are found to be heavily deformed. It may be stated that this work is suitable for practical applications.
Keywords: Post-processing; Arabic Numerals Recognition; Support Vector Machine; RBF Kernel.
A Pashtu Speakers Database using Accent and Dialect Approach
by Shahid Munir Shah, Shahzad Ahmed Memon, Khalil-ur-Rehman Khumbati, Muhammad Moinuddin
Abstract: In this research, a small scale Pashtu speakers database with different accents and dialects of Pashtu has been developed to use in Pashtu language based Speaker Identification systems (SIS) and Pashtu regional accents and dialect identification systems. Pashtu is a major spoken language of Pakistan and Afghanistan with its millions of speakers spread all over the world. For the past two decades, this language has been prominent due to its regional importance. The regions of Pakistan and Afghanistan where Pashtu language is spoken are mostly occupied by the extremists and they use Pashtu as their main mode of communication. Therefore, there exists a great need of Pashtu language based voice recognition systems to use in different security applications. As an initiative, here in this research a Pashtu speakers database has been presented. Due to rich nature of Pashtu language in accentual and dialectical variations, the database has been developed considering multiple accents and dialects of Pashtu. For this purpose, the data was recoded from Pashtu speakers of those regions of Pakistan and Afghanistan where Pashtu is spoken with different accents and dialects. The voice data in form of 25 small duration (1 to 3 seconds) daily conversational based sentences was collected from 32 native Pashtu speakers using different recording devices and smartphones. Finally, using a subset of the collected data, containing 5 voice samples of each speaker, a Mel Frequency Cepstral Coefficients (MFCC)-Multi-Layer Perceptron (MLP) based speaker identification system was designed to test the performance of the classifier with the collected data. The designed system achieved overall 87.5% recognition accuracy in identifying speakers based on their distinct accents and dialects. The comparative study shows that our MLP based system outperformed the recently proposed i-vector based and GMM based accent identification systems by showing 3.8% and 12.0% relative improvement, respectively, in recognition accuracy. In order to benchmark our investigation, a simple word recognition system using first 9 digits of Pashtu was also tested on MLP, HMM and SVM classifiers. SVM classifier, with 93.8 % recognition accuracy showed 1.3% and 3.3% relative improvement over MLP and HMM classifiers respectively, in recognizing Pashtu digits. Urdu accent of Pashtu was also tested in the study to check the performance of the proposed system on some other accent of (other than Pashtu). In case of Urdu accent of Pashtu, the system showed some decline in performance. It is because of the fact that the foreigners cannot pronounce Pashtu like the native Pashtu speakers. Finally, the robustness of our system against noise was investigated by varying SNR which shows that the performance of the proposed method slightly degrades with the increase in the noise level which reflects the robustness of the proposed system to noise.
Keywords: speaker recognition systems; accent and dialect identification systems; Pashtu speakers; Pashtu language.