Harvey A. Cohen

Classification and segmentation of rotated and scaled textured images using texture tuned masks

Abstract A rotation and scale invariant texture classifier function is described for effective classification and segmentation of images involving textures of unknown rotation and scale changes. The classifier used is the texture energy associated with a mask that has been “tuned” to be both discriminant between different textures, and to be invariant to rotation and scale changes. The mask tuning scheme utilized is based on task-oriented criterion optimization via a guided random search procedure to incorporate the changes. Both a dynamic texture sample set using a two-dimensional (2D) linked list and a re-ranking procedure are applied for training. Maximum feature dispersion of inter texture classes and high feature convergence of inner texture class samples associated with other statistical measures are suggested as key criteria in training. In a study based on 15 distinct Brodatz textures it is found that: the tuning process although computationally intensive converges efficiently; the mean classifier values of the classifier for a particular texture at different orientation and different scales are tightly clustered. An objective measure of classification capability is determined by computing the standard deviation of the classifier over pure texture at definite orientation and scale. Examples are presented of the classifier function applied to the segmentation of collages of Brodatz textures, comprising regions of various orientation and scale.

Efficient Image Matching

Image matching has played a key role in object recognition and localization. One central problem is to find an efficient and effective approach to search for the best matching between two image sets. In contrast to the conventional matching techniques, the innovation of our method detailed in this paper is to propose a hierarchical Chamfer matching scheme based on the dynamic detection of interesting points. The algorithm extends the traditional methods by introducing interesting points to replace edge points in distance transform for the matching measurement. The search for the best matching is guided by minimizing a given matching criterion in an interesting points pyramid from coarse level to fine level. The pyramid is created through a dynamic thresholding scheme and such a hierarchical structure aims to reduce the computation load. The processing speed is further improved by parallel implementation on a low cost heterogeneous PVM (Parallel Virtual Machine) network without specific software and hardware requirements. The experimental results demonstrate that our algorithm is simple to implement and quite insensitive to noise and other disturbances with reliability and efficiency.

Retrieval and Browsing of Images Using Image Thumbnails

In the emerging environment of massive image databases the display of a particular image at a workstation will be almost invariably preceded by its selection from a set of thumbnail images This paper explores the basics of image thumbnails, and the implications for efficient image storage. The advantages/disadvantages of alternate schemes for thumbnail generation are discussed. A measure of image compression is proposed to take into account the costs of multiple thumbs for each image accessed fully. Analysis suggests that to maximise the effective compression, requires the use of thumbs efficiently transferred that contain data required for whole image generation. To meet the needs of thumbnail-based retrieval, suggestions are made for the reorganization of the code for block-oriented image coding schemes leading to the specification and implementation of thumb-based vector quantization image coding, and a fast decoding thumb-fractal codec. A proposal for a thumb-oriented version of JPEG is outlined.

A hierarchical image matching scheme based on the dynamic detection of interesting points

This paper presents a parallel approach to a hierarchical image matching scheme using the Hausdorff distance for object recognition and localization in aerial images. Unlike the conventional matching methods in which edge pixels are considered as image feature pixels, the distance transform and the blind pointwise comparison procedure is simplified and extended in terms of the Hausdorff distance, and a guided image matching system is developed by the hierarchical detection of interesting points via a dynamic thresholding scheme for the search of the best matching between two image sets. Furthermore, the concept of remote procedure call (RPC) in distributed systems is introduced for the parallel implementation to achieve the speedup without specific software and hardware requirements.

Hierarchical image matching: a chamfer matching algorithm using interesting points

Image matching in conjunction with a distance transform has played an important role in computer vision and image analysis. This paper presents a new hierarchical chamfer matching algorithm based on the detection of interesting points. The algorithm extends the traditional method by introducing interesting points to replace edge points in the distance transform for the matching measurement. A series of images, with different numbers of interesting points featuring in the original image, is created in a pyramid structure through a dynamic thresholding scheme. The matching is performed in this pyramid in a coarse-to-fine level order, by minimizing a given matching criterion in terms of the distance between selected interesting points. This hierarchical structure aims to reduce the computational load. The algorithm is simple to implement and quite insensitive to noise and other disturbances. In addition, such a hierarchical matching scheme is implemented on a low-cost heterogeneous PVM (Parallel Virtual Machine) network to speed up the processing without any specific software and hardware requirements.

A multi-scale texture classifier based on multi-resolution “tuned” mask

Abstract Using a multi-resolution texture ‘tuned’ mask, we have realized a multi-scale texture classifier for effective segmentation of collages of Brodatz textures at various scales. Our mask tuning optimizes a performance index via a guided random search over a range of scales using a dynamic data structure.

An Efficient Parallel Texture Classification for Image Retrieval

This paper proposes an efficient parallel approach to texture classification for image retrieval. The idea behind this method is to pre-extract texture features in terms of texture energy measurement associated with a `tuned? mask and store them in a multiscale and multi-orientation texture class database via a two-dimensional linked list for query. Thus, each texture class sample in the database can be traced by its texture energy in a two-dimensional row-sorted matrix. The parallel searching strategies are introduced for fast identification of the entities closest to the input texture throughout the given texture energy matrix. In contrast to the traditional search methods, our approach incorporates different computation patterns for different cases of available processor numbers and concerns with robust and work-optimal parallel algorithms for row-search and minimum-find based on the accelerated cascading technique and the dynamic processor allocation scheme. Applications of the proposed parallel search and multisearch algorithms to both single image classification and multiple image classification are discussed. The time complexity analysis shows that our proposal will speed up the classification tasks in a simple but dynamic manner. Examples of the texture classification task applied to image retrieval of Brodatz textures, comprising various orientations and scales are presented.

Proposal for JPEG: thumbnail-based image access/retrieval

Image compression codecs should be adapted to the practical reality that most images are first viewed as thumbnail images, before the full size image is accessed. Existing JPEG variants, progressive and hierarchical, could be used for thumbnail based image access, but would involve a total re-engineering of existing libraries. An alternate proposal is made for a new variant of JPEG coding based on a reversible transformation of the JPEG output of existing compilers, so that existing decoder/encoders can be utilized, and existing JPEG libraries can be modified with no alteration to the quality of full size images after decompression. In the proposed scheme, the image code is partitioned into a thumb part and the remainder, or FF part. The thumb part is sufficient for the production of an image thumbnail, while this partition of the code together with the FF partition is required for full featured image reconstruction.

Deterministic scanning and hybrid algorithms for fast decoding of IFS (iterated function system) encoded image sets

Deterministic algorithms for decoding IFS (iterated function system) sets involve determining all the IFS (dynamic) descendants of seed pixels. Realistic algorithms require pruning of previously encountered pixels on the descendant tree. Timing data are reported for the random iteration algorithm, and for three new deterministic algorithms: the scanning algorithm; the stack algorithm; and a hybrid combination. Decoded timing data indicate the superiority of the pruned hybrid algorithm.<<ETX>>

Methods For Fast Image Object Recognition In Robotics

In Vision guided robotic operations, the rapid location and identification of image objects for assembly is essential. The image object data taken with a ccd camera may be in the form of a grey level, binary or colour image. This paper discusses our work on new speed up methods for object recognition in binary and grey level images. These are a coarse/fine stepping method and a sparse template technique. An extension of the coarse fine technique to give larger speedups is discussed. INTRODUCTION Searches for objects within images are extremely computationally expensive when scale, position and orientation of objects within the image is unknown. This paper reports results of a comprehensive coarse-fine object location program in which the criteria for coarse to fine switching have been investigated. These methods can speed up the column and row position search by a factor of 64. A matching error function is used to switch between coarse and fine search modes. Also a sparse template technique is used to give a further 64 fold speed up even more a for larger templates. This paper reports position search speed up times obtained on a number plate reading project. This work is of importance in vision guided assembly operations where machine vision techniques are used for locating parts and was first applied to a vehicle number plate location system. COARSE FINE COLUMN SEARCH By applying a reduced column search technique to object location in an image, a large reduction in computation may . be made. A basic approach is to move the template over the raster scan image, computing the mismatch function at each pixel location. Clearly if the matching calculations could be made at a reduced number of locations, ie a coarse search at say every fourth or fifth. pixel then a large speed up in the matching operation could be made. If a check is made by comparing the current matching error with the error calculation at the previous pixel position, then a decision can be made to move along say four or more columns, if the error is not moving towards a match. Alternatively, if the matching error is decreasing rapidly, from A to C in Fig 5, then the template pixel array will be moved along column by column after each calculation. Investigations showed that the template map may be moved along six columns for the coarse search and still find the matching position. A five to one speed up was obtained in this way. It is important to note that speedup factors are image dependent. COARSE FINE ROW SEARCH Similarly, a coarse fine (reduced) row search using a change in the image matching criterion between rows as a switch will give a speed up factor of the order of four or five to one assuming similar spatial changes in the vertical and horizontal directions of the image. This will be the case for images of uniform texture. More structured images of an industrial nature will not be so predictable. However matching error information is not as readily available in the case of the reduced row search. The matching error change method used was to save the lowest value of the matching error for the previous two rows and make a coarse fine search decision on the basis of this difference. An important point to consider in coarse fine searches is the dimensions of the template. A template four columns wide will have a much narrower correlation region than a 40 column wide template. Thus in certain cases for instance in the case of whole of image movement, where the template size is not fixed, it may be better to make the template longer (more columns) but not as wide, ie less rows if a coarse column search only is to be made. SPARSE TEMPLATES Matching error calculations may be speeded up by the use of sparse partial templates in which only say every fourth column and row brightness value of the template is checked against the image to be searched. A 16 times reduction in error calculations can be made in this way. For the case of noise-free images and templates, there is no need to use the full template. For the more practical case of differences between the image object and its template, the full template would be used when the matching error indicates that a match is close to the current search position. As the template becomes more sparse, the chance of falsely switching to a full template becomes quite large, setting an upper limit to the sparseness of the template. Another way the template calculation may be speeded up is by the use of a smaller template array. For textured or fractal type images, this is an important option. COARSE FINE SEARCH WITH BACKTRACKING. By allowing the column search to reverse in direction, the coarse-fine search step may be doubled, doubling the coarse fine speedup. Since the translation can reverse, we can move to the far side of the correlation region, point B in diagram 5 during the coarse search and then move backwards to the point of best match, ie point D. This gives a further factor of four speedup to the coarse-fine row, column search, APPLICATION TO OBJECTS OF VARYING SCALE If a plot is made of matching error versus image size or scale a finite correlation region typically of the order of 10 pixels will be found for all but completely random images. On the diagrams page, diagrams two and three 2 and 3 on the last page of the paper for the pebbles image. This result indicates that a finite set of templates may be used to find the object of unknown scale by making coarse steps in the template scale to form a finite template library. Again it is not necessary to perform a matching calculation over the entire template and sparse template techniques can also be used to reduce the number of matching calculations for each template. Sparse template techniques may also be used to reduce the calculation burden for each template. The number of competing template candidates in the image will a lso in f luence the template sca le s tep requi red to differentiate between the template objects. Also see reference 5.