Abstract To help farmers manage limited resources, rice disease diagnosis must be accurate, timely, and affordable. This study addresses challenges in rice field images, such as environmental variability and differences in rice leaf sizes. The proposed technique combines convolutional neural network object detection with image tiling, using estimated rice leaf width as a reference for image division. An 18‐layer ResNet model was trained using ground truth leaf width values for regression in leaf width estimation. Experiments used a dataset of 4960 images representing 8 rice diseases. The leaf width prediction model achieved a mean absolute percentage error of 11.18% and was used to generate a tiled dataset for training advanced rice disease detection models. The tiling technique was evaluated using YOLOv4, YOLOv8n, YOLOv8l, DINO‐5scale Swin‐L, and Co‐DINO‐5scale Swin‐L models by comparing detection performance on original and tiled datasets. Mean average precision improved significantly: YOLOv4 increased from 87.56% to91.14%, YOLOv8n from 89.80% to 91.70%, and YOLOv8l from 89.80% to 93.20%. More advanced models, such as DINO5scale Swin‐L and Co‐DINO‐5scale Swin‐L, achieved even higher precision, at 93.40% and 94.20%, respectively. In conclusion, the tiling technique improved detection efficiency and addressed object size variability, enhancing rice disease detection accuracy inreal‐world scenarios.

In this paper, we address the problem of land cover mapping where only the image labels are available in a training set. Here, a training sample is a pair of a remote sensing image and a class indicator of whether a land cover class is present or absent on the entire image. It is more economical to only identify what classes are present in an image than to label every pixel in an image. In this way, a large number of remote sensing images can be labeled quickly. To achieve this goal, we employed the attention layer to create the attention map. The attention map is then further segmented to produce the final land cover map where every pixel in an image will be labeled. We have tested the performance of our proposed algorithm with the UC Merced Dataset and achieved more than 85% overall accuracy with only 0.4% labels in a training set when comparing with the UNET [1].

Counterfactual inference has become a ubiquitous tool in online advertisement, recommendation systems, medical diagnosis, and econometrics. Accurate modelling of outcome distributions associated with different interventions---known as counterfactual distributions---is crucial for the success of these applications. In this work, we propose to model counterfactual distributions using a novel Hilbert space representation called counterfactual mean embedding (CME). The CME embeds the associated counterfactual distribution into a reproducing kernel Hilbert space (RKHS) endowed with a positive definite kernel, which allows us to perform causal inference over the entire landscape of the counterfactual distribution. Based on this representation, we propose a distributional treatment effect (DTE) which can quantify the causal effect over entire outcome distributions. Our approach is nonparametric as the CME can be estimated under the unconfoundedness assumption from observational data without requiring any parametric assumption about the underlying distributions. We also establish a rate of convergence of the proposed estimator which depends on the smoothness of the conditional mean and the Radon-Nikodym derivative of the underlying marginal distributions. Furthermore, our framework allows for more complex outcomes such as images, sequences, and graphs. Our experimental results on synthetic data and off-policy evaluation tasks demonstrate the advantages of the proposed estimator.

This paper presents a study of two simple methods for reducing the complexity of the instance-based classification technique and demonstrates their use in device-context independent activity recognition on a mobile phone. A projection-based method for signal rectification has been implemented on an iPhone in order to handle with variation in device orientations. The transformation matrix is estimated on a ten-second dynamic data buffer. To search for a suitable set of training prototypes for iPhone implementation, an activity recognition experiment is conducted with twenty different device contexts performed by eight subjects. With the developed mobile application, the recognition results along with the user's location can be displayed on both iPhone and the web application in real time.

This paper investigates two major issues in using a tri-axial accelerometer-embedded mobile phone for continuous activity monitoring, i.e. the difference in orientations and locations of the device. Two experiments with a total of ten test subjects performed six daily activities were conducted in this study: one with a device fixed on the waist in sixteen different orientations and another with three different device locations (i.e., shirt-pocket, trouser-pocket and waist) in two different device orientations. For handling with varying device orientations, a projection-based method for device coordinate system estimation has been proposed. Based on the dataset with sixteen different device orientations, the experimental results have illustrated that the proposed method is efficient for rectifying the acceleration signals into the same coordinate system, yielding significantly improved activity recognition accuracy. After signal transformation, the recognition results of signals acquired from different device locations are compared. The experimental results show that when the sensor is placed on different rigid body, different models are required for certain activities.

We investigate a new approach for online handwritten shape recognition. Interesting features of this approach include learning without manual tuning, learning from very few training samples, incremental learning of characters, and adaptation to the user-specific needs. The proposed system can deal with two-dimensional graphical shapes such as Latin and Asian characters, command gestures, symbols, small drawings, and geometric shapes. It can be used as a building block for a series of recognition tasks with many applications.

In this paper, we proposed a new two-dimensional linear discriminant analysis (2DLDA) method. Based on two-dimensional principle component analysis (2DPCA), face image matrices do not need to be previously transformed into a vector. In this way, the spatial information can be preserved. Moreover, the 2DLDA also allows avoiding the small sample size (SSS) problem, thus overcoming the traditional LDA. We combine 2DPCA and our proposed 2DLDA on the two-dimensional linear discriminant analysis of principle component vectors framework. Our framework consists of two steps: first we project an input face image into the family of projected vectors via 2DPCA-based technique, second we project from these space into the classification space via 2DLDA-based technique. This does not only allows further reducing of the dimension of the feature matrix but also improving the classification accuracy. Experimental results on ORL and Yale face database showed an improvement of 2DPCA-based technique over the conventional PCA technique

This paper proposes an off-line Thai character recognition system for an unconstrained writing. This system combines both global feature, that represents character shape, and local feature, that represents the symbolic structure, to solve the similarity and variety problems of Thai characters. The pixel distribution is applied, as the global feature, to classify the image into 20 groups of similar characters. The modified structure features extraction is used to extract character structure. Then, all features, global and local, are used to classify the character within each group. The proposed method yields the recognition rate at 87.32 per cents.

A major feature of new mobiles terminals using penbased interfaces, such as personal assistants or e-book, is their personal character, implying that a good interface should be easily customizable in order to meet various users ’ needs. We proposed recently a new recognition engine with strong adaptation abilities that allows learning a user’s writing style or new symbols easily. It dealt with characters with simple shape only; we describe here an extension of this recognition system that deals with more complex graphical symbols. We propose an adaptive learning scheme that can learn more and more as the writer gives new samples. 1.

This paper focuses on designing a handwriting recognition system dealing with on-line signal, i.e. temporal handwriting signal captured through an electronic pen or a digitalized tablet. We present here some new results concerning a hybrid on-line handwriting recognition system based on Hidden Markov Models (HMMs) and Neural Networks (NNs), which has already been presented in several contributions. In our approach, a letter-model is a Left-Right HMM, whose emission probability densities are approximated with mixtures of predictive multilayer perceptrons. The basic letter models are cascaded in order to build models for words and sentences. At the word level, recognition is performed thanks to a dictionary organized with a tree-structure. At the sentence level, a word-predecessor conditioned frame synchronous beam search algorithm allows to perform simultaneously segmentation into words and word recognition. It processes through the building of a word graph from which a set of candidate sentences may be extracted. Word and sentence recognition performances are evaluated on parts of the UNIPEN international database.