Hemerson Pistori

Mice and larvae tracking using a particle filter with an auto-adjustable observation model

This paper proposes a novel way to combine different observation models in a particle filter framework. This, so called, auto-adjustable observation model, enhance the particle filter accuracy when the tracked objects overlap without infringing a great runtime penalty to the whole tracking system. The approach has been tested under two important real world situations related to animal behavior: mice and larvae tracking. The proposal was compared to some state-of-art approaches and the results show, under the datasets tested, that a good trade-off between accuracy and runtime can be achieved using an auto-adjustable observation model.

Multiple Mice Tracking using a Combination of Particle Filter and K-Means

This paper presents a data hiding technique for printed bicolor documents. It inserts tiny dots, hardly noticeable at normal reading distance, to embed the message. For message extraction, we employ autocorrelation and tiny registration dots to rectify geometric distortions. This technique is robust to distortions resulting from print-scan operations, good quality photocopies, affine transformations and scribblings/stains on the paper. The technique can be applied to documents with large white (or black) areas and they may present characters, drawings, schematics, diagrams, cartoons, but not halftones. The technique is intended to be neither a robust watermark (because any filtering can remove the dots) nor a covert communication (because the dots are perceptible at short distance). Nevertheless, when combined with a perceptual hashing and a cryptography protocol, it can be applied as semi-fragile authentication watermarking for hardcopy two-tone documents. In some situations, the utilization of the proposed system can substitute the use of notarial authenticated photocopies.This paper presents a new approach to multiple objects tracking that combines particle filters and k-means. The approach has been tested under an important real world situation, related to pharmacological development, which has also proved to serve as an interesting ground-truth dataset provider for the evaluation of tracking algorithms. The obtained results are then compared to other models. The promising results of these experiments are presented.

Adaptive Finite State Automata and Genetic Algorithms: Merging Individual Adaptation and Population Evolution

This paper presents adaptive finite state automata as an alternative formalism to model individuals in a genetic algorithm environment. Adaptive finite automata, which are basically finite state automata that can change their internal structures during operation, have proven to be an attractive way to represent simple learning strategies. We argue that the merging of adaptive finite state automata and GA results in an elegant and appropriate environment to explore the impact of individual adaptation, during lifetime, on population evolution.

An experiment on handshape Sign recognition using adaptive technology: Preliminary results

This paper presents an overview of current work on the recognition of sign language and a prototype of a simple editor for a small subset of the Brazilian Sign Language, LIBRAS. Handshape based alphabetical signs, are captured by a single digital camera, processed on-line by using computational vision techniques and converted to the corresponding Latin letter. The development of such prototype employed a machine-learning technique, based on automata theory and adaptive devices. This technique represents a new approach to be used in the far more complex problem of full LIBRAS recognition. As it happens with spoken languages, sign languages are not universal. They vary a lot from country to country, and in spite of the existence of many works in American Sign Language (ASL), the automatic recognition of Brazilian Sign Language has not been extensively studied. ...

Decision Tree Induction using Adaptive FSA

This paper introduces a new algorithm for the induction of decision trees, based on adaptive techniques. One of the main feature of this algorithm is the application of automata theory to formalize the problem of decision tree induction and the use of a hybrid approach, which integrates both syntactical and statistical strategies. Some experimental results are also presented indicating that the adaptive approach is useful in the construction of ecien t learning algorithms.

Weed detection in soybean crops using ConvNets

Abstract Weeds are undesirable plants that grow in agricultural crops, such as soybean crops, competing for elements such as sunlight and water, causing losses to crop yields. The objective of this work was to use Convolutional Neural Networks (ConvNets or CNNs) to perform weed detection in soybean crop images and classify these weeds among grass and broadleaf, aiming to apply the specific herbicide to weed detected. For this purpose, a soybean plantation was carried out in Campo Grande, Mato Grosso do Sul, Brazil, and the Phantom DJI 3 Professional drone was used to capture a large number of crop images. With these photographs, an image database was created containing over fifteen thousand images of the soil, soybean, broadleaf and grass weeds. The Convolutional Neural Networks used in this work represent a Deep Learning architecture that has achieved remarkable success in image recognition. For the training of Neural Network the CaffeNet architecture was used. Available in Caffe software, it consists of a replication of the well known AlexNet, network which won the ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC2012). A software was also developed, Pynovisao, which through the use of the superpixel segmentation algorithm SLIC, was used to build a robust image dataset and classify images using the model trained by Caffe software. In order to compare the results of ConvNets, Support Vector Machines, AdaBoost and Random Forests were used in conjunction with a collection of shape, color and texture feature extraction techniques. As a result, this work achieved above 98% accuracy using ConvNets in the detection of broadleaf and grass weeds in relation to soil and soybean, with an accuracy average between all images above 99%.

A new strategy for applying grammatical inference to image classification problems

This paper presents a new strategy to represent an image as a string so that standard grammar induction techniques can be used in computer vision problems. Two sets of experiments using an artificial and a real dataset have been conducted in order to explore the new strategy parameters and to have a first glimpse on its comparative performance against some standard machine learning techniques. The results are encouraging and the proposal opens new paths of exploration for syntactical pattern recognition.

Hidden Markov models applied to snakes behavior identification

This paper presents an application of the hidden Markov models (HMMs) to the recognition of snakes behaviors, an important and hard problem that, as far as the authors know, has not been tackled before, by the computer vision community. Experiments were conducted using different HMM configurations, including modifications on the number of internal states and the initialization procedures. The best results have shown a 84% correct classification rate, using HMMs with 4 states and an initialization procedure based on the K-Means algorithm.

SVM with stochastic parameter selection for bovine leather defect classification

The performance of Support Vector Machines, as many other machine learning algorithms, is very sensitive to parameter tuning,mainly in real world problems. In this paper, two well known and widely used SVM implementations, Weka SMO and LIBSVM, were compared using Simulated Annealing as a parameter tuner. This approach increased significantly the classification accuracy over the Weka SMO and LIBSVM standard configuration. The paper also presents an empirical evaluation of SVM against AdaBoost and MLP, for solving the leather defect classification problem. The results obtained are very promising in successfully discriminating leather defects, with the highest overall accuracy, of 99.59%, being achieved by LIBSVM tuned with Simulated Annealing.

Combining color and haar wavelet responses for aerial image classification

A new set of attributes combining color and SURF-based histograms coupled with a SVM classifier to enhance visual based autonomous aerial navigation is proposed. These new features are used for region classification with aerial images in order to speed up the UAV (Unmanned Aerial Vehicles) localization performed by image matching using only reference images according to the region classification. Experimental results comparing the proposal with color or SURF only attributes are presented. In the experiments the UAV localization task can be performed four times faster using the proposed approach, however the performance gain can be still bigger for large datasets of reference images.