Capturing multi-scale context within feature maps is crucial for semantic segmentation. With the success of the Vision Transformer (ViT), recent models have been designed with transformer decoders to capture it. However, these models face limitations in utilizing diverse contextual information due to the inherent nature of the attention mechanism and structural constraints. Typically, multi-head attention which leads to similar receptive fields for each token feature is achieved at the expense of significantly increased computational cost. The nature of the structure can cause inconsistent combination of the information across different levels. To address this issue, in this paper, we propose a novel and effective decoding scheme, CCASeg, which is based on convolutional cross-attention (CCA). The proposed CCA along with the decoding structure is devised not only to capture both local and global context through convolutional kernels of various sizes, but also to achieve high efficiency by effective utilization of the cheap convolution operations. Moreover, the decoding structure, which ensures the successive combination of information across various levels, facilitates understanding of diverse contexts. Consequently, this novel decoding scheme enables feature maps to effectively learn the relationships between objects of different sizes. In this way, our proposed CCASeg outperforms previous state-of-the-art methods on popular semantic segmentation benchmarks, including ADE20K, Cityscapes, COCO-stuff, and iSAID.

Multi-scale architecture, including hierarchical vision transformer, has been commonly applied to high-resolution semantic segmentation to deal with computational complexity with minimum performance loss. In this paper, we propose a novel decoding scheme for semantic segmentation in this regard, which takes multi-level features from the encoder with multi-scale architecture. The decoding scheme based on a multi-level vision transformer aims to achieve not only reduced computational expense but also higher segmentation accuracy, by introducing successive cross-attention in aggregation of the multi-level features. Furthermore, a way to enhance the multilevel features by the aggregated semantics is proposed. The effort is focused on maintaining the contextual consistency from the perspective of attention allocation and brings improved performance with significantly lower computational cost. Set of experiments on popular datasets demonstrates superiority of the proposed scheme to the state-of-the-art semantic segmentation models in terms of computational cost without loss of accuracy, and extensive ablation studies prove the effectiveness of ideas proposed.

Recently, there have been many performance improvements in super resolution using deep learning methods. However, despite the performance improvement, it remains a challenge to reduce the amount of computation for real products. In this letter, we propose a deep network that employs modified Squeezenet's fire modules. We introduce a way to modify the original fire module for effective separation of spatial- and channel-wise learning, and describe how the modified fire modules can be arranged asymmetrically for reducing the number of parameters of the network. Our experiment results show higher PSNR and competitive processing time to other super resolution networks, but with less number of parameters.

In this letter, we propose a single image haze removal method, which is inherently an ill-posed problem due to the transmission estimation that is dependent on the depth information. The pointwise haze density estimation method using conditional random fields includes the unary factor that encodes the likelihood for the haze density and the pairwise factor that encodes the spatial contextual relationship between adjacent unary factors. The proposed scheme makes elimination of the refining process possible and it may prevent the halo effect at object boundaries. The results of experiments carried out on various real and synthesized images, and comparison with existing methods are presented to verify the effectiveness of the proposed method.

Blur kernel estimation is a crucial step in the deblurring process for images. Estimation of the kernel, especially in the presence of noise, is easily perturbed, and the quality of the resulting deblurred images is hence degraded. Since every motion blur in a single exposure image can be represented by 2D parametric curves, we adopt a piecewise-linear model to approximate the curves for the reliable blur kernel estimation. The model is found to be an effective tradeoff between flexibility and robustness as it takes advantage of two extremes: (1) the generic model, represented by a discrete 2D function, which has a high degree of freedom (DOF) for the maximum flexibility but suffers from noise and (2) the linear model, which enhances robustness and simplicity but has limited expressiveness due to its low DOF. We evaluate several deblurring methods based on not only the generic model, but also the piecewise-linear model as an alternative. After analyzing the experiment results using real-world images with significant levels of noise and a benchmark data set, we conclude that the proposed model is not only robust with respect to noise, but also flexible in dealing with various types of blur.

An approach for segmenting handwritten text in a pre-hospital care report (PCR) is presented. Segmentation of lines and words in a PCR is extremely challenging due to the nature of the environment in which the reports are created, giving rise to low quality, poorly written, loosely constrained data. Stroke analyses are performed and image primitives are extracted for word detection. A heuristics-based approach, involving gap spacing, height transitions, and the average stroke width of the writer is used in detecting word boundaries. Carbon copies of live PCRs are used for testing. Experiments show perfect segmentation of 69%, outperforming the more tested and proven algorithms by as much as 15%.

An approach for segmentation of handwritten touching numeral strings is presented in this paper. A neural network has been designed to deal with various types of touching observed frequently in numeral strings. A numeral string image is split into a number of line segments while stroke extraction is being performed and the segments are represented with straight lines. Four types of primitive are defined based on the lines and used for representing the numeral string in more abstractive way and extracting clues on touching information from the string. Potential segmentation points are located using the neural network by active interpretation of the features collected from the primitives. Also, the run-length coding scheme is employed for efficient representation and manipulation of images. On a test set collected from real mail pieces, the segmentation accuracy of 89.1% was achieved, in image level, in a preliminary experiment.

A systematic approach for encoding Korean addresses to the finest depth of sort is presented in this paper. The implementation is focused on producing the final delivery point code for various types of address recognized in an efficient manner. There are two stages in the address interpretation: 1) agreement verification between the recognized postal code and upperpart of the address; and 2) analysis of the lower part of the address which is important for the encoding. In the agreement verification procedure, the recognized postal code is used as a key to access the address dictionary and each of the retrieved addresses is compared with the words in the recognized address. As a result, the boundary between the upper part and the lower part is located. The confusion matrices are introduced to improve the performance of the process by correcting misrecognized characters. In the procedure of interpreting the lower address part, a delivery point code is derived using the house number and/or the building name. Several rules for the interpretation have been developed based on the analysis of real addresses collected. Experiments have been performed to evaluate the proposed approach using addresses collected from two metropolitan cities in Korea.

String recognition is rather paradoxical problem because it requires the segmentation of the string into understandable units, but proper segmentation needs a-priori knowledge of the units and this implies a recognition capability. To solve this dilemma therefore, both a-priori knowledge of meaningful units and a segmentation method have to be used together, and they should dynamically interact with each other. In other words, the results of segmentation are used as fundamental information to suppose what is most likely to be, and then its a-priori knowledge is used to help the segmentation. This model makes explicit segmentation unnecessary because it does not speculate on possible break positions. It is also possible to recognize a digit even if it contains strokes that do not belong to to it. Using this paradigm for 100 handwritten numeral strings belonging to the NIST database has resulted in 95% recognition.

Contour representations of binary images of handwritten words afford considerable reduction in storage requirements while providing lossless representation. On the other hand, the one-dimensional nature of contours presents interesting challenges for processing images for handwritten word recognition. Our experiments indicate that significant gains are to be realized in both speed and recognition accuracy by using a contour representation in handwriting applications.

A bankcheck reading system using cross validation of both the legal and the courtesy amounts is presented in this paper. Some of the challenges posed by the task are (i) segmentation of the legal amount into words, (ii) location of boundaries between dollars and cents amounts, and (iii) high accuracy in terms of recognition performance. Word segmentation in the legal amount is a serious issue because of the nature of the data and patrons' writing habits which tend to clump words together. We have developed a word segmentation algorithm based on the character segmentation results to address this issue. The list of possible amounts generated by the word segmentation hypotheses is used as lexicon for the courtesy amount recognition. The order of magnitude of the amount is estimated during legal amount recognition. We treat the courtesy amount as a numeral string and apply the same word recognition scheme as used for the legal amount. Our approach to check recognition differs from traditional methods in two significant aspects: First, our emphasis on both the legal and the courtesy amounts is balanced. We use an accurate word recognizer which performs equally well on alpha words and digit strings. Second, our combination strategy is serial rather than the commonly used parallel method. Experimental results show that 43.8% of check images are correctly read with an error rate of 0%.

A fast method of handwritten word recognition suitable for real time applications is presented in this paper. Preprocessing, segmentation and feature extraction are implemented using a chain code representation of the word contour. Dynamic matching between characters of a lexicon entry and segment(s) of the input word image is used to rank the lexicon entries in order of best match. Variable duration for each character is defined and used during the matching. Experimental results prove that our approach using the variable duration outperforms the method using fixed duration in terms of both accuracy and speed. Speed of the entire recognition process is about 200 msec on a single SPARC-10 platform and the recognition accuracy is 96.8 percent are achieved for lexicon size of 10, on a database of postal words captured at 212 dpi.

A method for recognition of street name phrases collected from mail pieces is presented in this paper. Some of the challenges posed by the problem are: (i) patron errors, (ii) non-standardized way of abbreviating names, and (iii) variable number of words in a street name image. A neural network has been designed to segment words in a phrase, a street name in this case, using distances between components and style of writing. The network learns the type of spacing (including size) that one should expect between different pairs of characters in handwritten test. Experiments show perfect word segmentation performance at about 85% of cases. Unlike conventional methods, where lexicon entries are expanded to take care of all variations of prefixes and sizes, substring matching is attempted only between the main body of a lexicon entry and the word segments of an image. Efforts to reduce computational complexity are successfully made by the sharing of character segmentation results between the segmentation and recognition phases. 83% phrase recognition accuracy is achieved on a test set.

A fast matching algorithm using dynamic programming for handwritten word recognition is presented. The variable duration for each character is defined based on the character segmentation statistics obtained during the training phase. The variable duration represents the possible ways in which a character image can be split into segments. Experimental results prove that our approach using the variable duration outperforms the method using fixed duration in terms of both accuracy and speed. The processing time for the matching process increases at the rate of log(N) as the lexicon size N increases. A speed of 200 msec, including image handling procedures, on a single Sparc 10 platform and a recognition accuracy of 96.7% are achieved for a lexicon size of 10.

A segmentation and recognition method for handwritten phrases, such as street names, is presented in this paper. Some of the challenges posed by the problem are: (1) identifying correct word gaps from character gaps and (2) minimization of computational complexity during the recognition of potential words. A trainable word segmentation scheme using a neural network is introduced. The network learns the type of spacing (including size) that one should expect between different pairs of characters in handwritten text. The concept of variable duration, which is obtained during the training phase of a word recognition engine we have developed, is expanded to reduce the computational complexity which has been a serious concern in this type of application.

A fast handwritten word recognition system for real time applications is presented. Preprocessing, segmentation and feature extraction are implemented using chain code representation. Dynamic matching between each character of a lexicon entry and segment(s) of input word image is used for ranking words in the lexicon. Speed of the entire recognition process is about 200 msec on a single SPARC-10 platform for lexicon size of 10. A top choice performance of 96% is achieved on a database of postal words captured at 212 dpi.