Are there approaches other than convolutions to learn features from images? Has there been any research to use approaches such as hashing (e.g.
diff-hash etc.) in lieu?
Some more complex and better outcomes are those that include the MPEG-7 developed by the MPEG group.
These descriptors specify the video standard in question. They are split for color, shape, texture and movement.
1. Color Descriptors
Color Space. Specifies the data type of the color space in which they are expressed or work other color descriptors. Color spaces that it contemplated are: RGB (Red, Green Blue), YCrCb (Luminance, Chrominance), HSV (Hue, Saturation, Value), HMMD (Hue, Maximum, Minimum, and Difference).
Color Quantization. This descriptor defines a uniform quantization of a given color space. Dominant Color(s). This descriptor is the most suitable to be used in images, or regions, in which a small number of colors is sufficient to characterize the information of one determined region.
Scalable Color. This descriptor consists in a color histogram in the HSV space. It is useful in image to image comparisons or searches based on color characteristics.
Color Layout. This descriptor allows representing the color spatial distribution within the images in a very compact way, so the recovery is realized with great efficiency.
Color Structure Descriptor. This descriptor characterizes the color distribution in an image. It builds a color histogram, which assigns most importance to the colors that appears most often and spread across the image. GoF/GoP Color (Group of Frames/Group of Pictures). This descriptor is an extension of the Scalable Color descriptor, which, unlike the latter, is applied to video sequences (a collection of images) instead of only an image.
2. Texture Descriptors
Homogeneous Texture. This descriptor emerged as an important tool when looking for and choosing within large collections of images of great visual similarity.
Texture Browsing. This descriptor specifies the perceptual characterization of one particular texture, and pretend be similar to the characterization that a human eye makes, according to regularity terms, coarseness and directivity. Edge Histogram. It is a descriptor that gives us information on the type of contours or edges that appear in the image.
3. Shape Descriptors
Region Shape. The object´s shape in an image may consist of a single region or a set of regions, so this descriptor is useful for this type of characterization. Contour Shape. It is characterized by very well represented contour features which facilitates subsequent search and retrieval; it is robust to motion, to occlusions and to different perspectives; it is extremely compact.
Shape 3D. The Shape 3D allows to describe in details the shape in 3D. This tool is very useful today due to continuous development of multimedia technologies.
This was taked from my paper: The MPEG-7 Visual Descriptors: a Basic Survey
I think your question addresses the fact that convolutional neural networks utilize convolution and not some other mechanism. This answer addresses other feature extraction techniques you can use on image. In general, there are a number of other feature extraction mechanisms you can look into.
- Histogram oriented gradients. Counts the occurrences of gradients in an image and generates bins. These bins can be then used directly as features.
- Color data. You can extract color histograms of various granularities and use this data to train a model.
- Hough Transform - Can identify shapes and lines and other spatial features.
On top of these we also have numerous techniques on top of convolutions like gradients, curvature finding, corner detection, and ridge finding.