Template matching

The hidden layer outputs a vector that holds classification information about the image and is used in the Template Matching algorithm as the features of the image

Feature-based approach relies on the extraction of image features such, i.e. shapes, textures, colors, to match in the target image or frame. This approach is currently achieved by using Neural Networks and Deep Learning classifiers such as VGG,[6] AlexNet, ResNet. Deep Convolutional Neural Networks process the image by passing it through different hidden layers and at each layer produce a vector with classification information about the image. These vectors are extracted from the network and are used as the features of the image. Feature extraction by using Deep Neural Networks is extremely effective and thus is the standard in state of the art template matching algorithms.[7]

This method is considered more robust and is state of the art as it can match templates with non-rigid and out of plane transformation, it can match with high background clutter and illumination changes.[8][9] [10]

For templates without strong features, or for when the bulk of the template image constitutes the matching image, a template-based approach may be effective. As aforementioned, since template-based matching may potentially require sampling of a large number of points, it is possible to reduce the number of sampling points by reducing the resolution of the search and template images by the same factor and performing the operation on the resultant downsized images (multiresolution, or pyramid), providing a search window of data points within the search image so that the template does not have to search every viable data point, or a combination of both.

In instances where the template may not provide a direct match, it may be useful to implement the use of eigenspaces – templates that detail the matching object under a number of different conditions, such as varying perspectives, illuminations, color contrasts, or acceptable matching object “poses”.[11] For example, if the user was looking for a face, the eigenspaces may consist of images (templates) of faces in different positions to the camera, in different lighting conditions, or with different expressions.

It is also possible for the matching image to be obscured, or occluded by an object; in these cases, it is unreasonable to provide a multitude of templates to cover each possible occlusion. For example, the search image may be a playing card, and in some of the search images, the card is obscured by the fingers of someone holding the card, or by another card on top of it, or any object in front of the camera for that matter. In cases where the object is malleable or poseable, motion also becomes a problem, and problems involving both motion and occlusion become ambiguous.[12] In these cases, one possible solution is to divide the template image into multiple sub-images and perform matching on each subdivision.

Template matching is a central tool in Computational anatomy (CA). The deformable template model[13] models the space of human anatomies and orbits under the group action of diffeomorphisms. Template matching arise as a problem in matching the unknown diffeomorphism that acts on the template to match the target image.

Template matching algorithms in CA have come to be called large deformation diffeomorphic metric mapping (LDDMM); there are now LDDMM template matching algorithms for matching landmark points, curves, surfaces, volumes.

A basic method of template matching uses an image patch (template), tailored to a specific feature of the search image, which we want to detect. This technique can be easily performed on grey images or edge images. The cross correlation output will be highest at places where the image structure matches the mask structure, where large image values get multiplied by large mask values.

This method is normally implemented by first picking out a part of the search image to use as a template: We will call the search image S(x, y), where (x, y) represent the coordinates of each pixel in the search image. We will call the template T(x _t, y _t), where (x_t, y_t) represent the coordinates of each pixel in the template. We then simply move the center (or the origin) of the template T(x _t, y _t) over each (x, y) point in the search image and calculate the sum of products between the coefficients in S(x, y) and T(x_t, y_t) over the whole area spanned by the template. As all possible positions of the template with respect to the search image are considered, the position with the highest score is the best position. This method is sometimes referred to as 'Linear Spatial Filtering' and the template is called a filter mask .

Another way to handle translation problems on images using template matching is to compare the intensities of the pixels, using the SAD (Sum of absolute differences) measure.

A pixel in the search image with coordinates (x_s, y_s) has intensity I_s(x_s, y_s) and a pixel in the template with coordinates (x_t, y_t) has intensity I_t(x_t, y_t ). Thus the absolute difference in the pixel intensities is defined as Diff(x_s, y_s, x _t, y _t) = | I_s(x_s, y_s) – I_t(x _t, y _t) |.

${\displaystyle SAD(x,y)=\sum _{i=0}^{T_{\text{rows}}}\sum _{j=0}^{T_{\text{cols}}}{{\text{Diff}}(x+i,y+j,i,j)}}$

The mathematical representation of the idea about looping through the pixels in the search image as we translate the origin of the template at every pixel and take the SAD measure is the following:

${\displaystyle \sum _{x=0}^{S_{\text{rows}}}\sum _{y=0}^{S_{\text{cols}}}{SAD(x,y)}}$

S_rows and S_cols denote the rows and the columns of the search image and T_rows and T_cols denote the rows and the columns of the template image, respectively. In this method the lowest SAD score gives the estimate for the best position of template within the search image. The method is simple to implement and understand, but it is one of the slowest methods.

In this simple implementation, it is assumed that the above described method is applied on grey images: This is why Grey is used as pixel intensity. The final position in this implementation gives the top left location for where the template image best matches the search image.

minSAD = VALUE_MAX;

// loop through the search image
for ( size_t x = 0; x <= S_cols - T_cols; x++ ) {
    for ( size_t y = 0; y <= S_rows - T_rows; y++ ) {
        SAD = 0.0;

        // loop through the template image
        for ( size_t j = 0; j < T_cols; j++ )
            for ( size_t i = 0; i < T_rows; i++ ) {

                pixel p_SearchIMG = S[y+i][x+j];
                pixel p_TemplateIMG = T[i][j];
		
                SAD += abs( p_SearchIMG.Grey - p_TemplateIMG.Grey );
            }

        // save the best found position 
        if ( minSAD > SAD ) { 
            minSAD = SAD;
            // give me min SAD
            position.bestRow = y;
            position.bestCol = x;
            position.bestSAD = SAD;
        }
    }
    
}

One way to perform template matching on color images is to decompose the pixels into their color components and measure the quality of match between the color template and search image using the sum of the SAD computed for each color separately.

In the past, this type of spatial filtering was normally only used in dedicated hardware solutions because of the computational complexity of the operation,[14] however we can lessen this complexity by filtering it in the frequency domain of the image, referred to as 'frequency domain filtering,' this is done through the use of the convolution theorem.

Another way of speeding up the matching process is through the use of an image pyramid. This is a series of images, at different scales, which are formed by repeatedly filtering and subsampling the original image in order to generate a sequence of reduced resolution images.[15] These lower resolution images can then be searched for the template (with a similarly reduced resolution), in order to yield possible start positions for searching at the larger scales. The larger images can then be searched in a small window around the start position to find the best template location.

Other methods can handle problems such as translation, scale, image rotation and even all affine transformations.[16][17][18]

Improvements can be made to the matching method by using more than one template (eigenspaces), these other templates can have different scales and rotations.

It is also possible to improve the accuracy of the matching method by hybridizing the feature-based and template-based approaches.[19] Naturally, this requires that the search and template images have features that are apparent enough to support feature matching.

Other methods which are similar include 'Stereo matching', 'Image registration' and 'Scale-invariant feature transform'.

Template matching has various applications and is used in such fields as face recognition (see facial recognition system) and medical image processing. Systems have been developed and used in the past to count the number of faces that walk across part of a bridge within a certain amount of time. Other systems include automated calcified nodule detection within digital chest X-rays.[20] Recently, this method was implemented in geostatistical simulation which could provide a fast algorithm.[21]

• Facial recognition system
• Pattern recognition
• Computer vision
• Elastic Matching

• Template Matching in OpenCV
• Visual Object Recognition using Template Matching
• Rotation, scale, translation-invariant template matching demonstration program
• perspective-invariant template matching