1, convolution
A specific example is given below: Suppose you have learned from the image of a 96x96 to 8x8 samples it has a feature, which is assumed by the self-encoded with 100 hidden units is completed. In order to obtain convolution characteristics required for each of the small image area of 96x96 8x8 convolution operation are performed. That is, the 8x8 small block area extracted, and sequentially labeled from a start coordinate (1,1), (1,2), ..., up to (89, 89), and then individually extracted region running from trained sparse coding to get the activation value characteristics. In this example, will be apparent to give 100, each set containing 89x89 convolution characteristics.
As shown below, illustrates the process the 3 × 3 convolution kernel to do the convolution on the image of 5 × 5. Each convolution is a characteristic extraction method, like a sieve, a portion of the condition (activated larger values meet the conditions) that match an image filtered out.
2: that under the pooling, pooling in fact very easy to understand, look at the chart:
Transfer: http://blog.csdn.net/silence1214/article/details/11809947
Such as upper left matrix A is a matrix to be 20 * 20 10 * the size of the pool 10, then the red left in the figure is the size of 10 * 10, which corresponds to the right side of the matrix, each element of the right value, is a left side of each element is worth matrix red and red again the number of elements in the matrix, i.e. in the form of an average value of the pool.
3: The above said lower convolution and pooling, say the next calculation need to be aware of. Code is used in the color image, color image has three channels, each channel is then to be convoluted, and pooled separately, especially when there is a place to be noted convoluting, for each hidden layer a value corresponding to one of the three channels of FIG wear, it is divided into three channels is performed after convolution to add up to exactly correspond to the neurons in a hidden layer, i.e. up a feature.
On a sample image (upsampling) and downsampling (subsampled)
The main object of the reduced image (or downsampling (subsampled) or downsampling (downsampled)) are two:
1, so that the image fit the size of the display region; 2, generating a thumbnail image corresponding to.
The main object of the enlarged image (or upsampling (upsampling) or image interpolation (Interpolating)) are:
Enlarged original image can be displayed on a higher resolution display apparatus. Image scaling operation does not lead to more information about the image, the image quality will inevitably be affected. However, the method does have some zoom information of the image can be increased, so that the quality of the scaled image than the original image quality.
The sampling theorem: For an image I of size M * N, the sample subjected to the s-fold, i.e., to give (M / s) * to obtain resolution images (N / s) size, of course, M and N should be s the divisor for the job, if the consideration is in the form of a matrix image is the image in the original image s * s window into a pixel value is the average of all the pixels within the window of the pixel point:
The sampling theory: almost all enlarged image using the interpolation method, i.e. using a suitable interpolation algorithm to insert a new element in the pixels between the original pixels on the image.
Whether the image scaling (down-sampled) image is enlarged (upsampled), there are a variety of sampling methods. The nearest neighbor interpolation, bilinear interpolation, interpolation of the mean, median interpolation method. In AlexNet on the use of a more suitable interpolation method. Various interpolation method has its own advantages and disadvantages.