Target detection
First, the definition of target detection:
Target detection is in the picture variable number to find and classify targets
Second, the target detection process common problems:
- Target species and the number of questions
- Target scale problems
- The external environment interference
Third, the object detection VS image classification:
Target Detection:
Target detection is not only given a rectangular frame (target position detection object), while the rectangle object classified in a rectangular frame different colors represent different classes, and gives the detection of an object belonging to the target class of KAP
Image Category:
Image classification mainly as an input image, the image belong to different categories of probability distributions as output, mainly for image category determination
Whether image classification or target detection, when using deep learning technologies for processing, we need to feature extraction part, for classical machine learning methods, often through design manual features, to complete the feature extraction, and deep learning often by volume neural network for completing the product extracted features
Fourth, the target detection object segmentation VS
- Image Classification: category just specify the corresponding target belongs
- Object detection: the need to locate the position of the target is located, and classifies
- Object segmentation: the need to find the current target region occupied by the divided semantic FIG c, d is an example of division
- Semantic segmentation: just need to find the area occupied by the same type of target
- Examples divided: not only to distinguish between different semantic level of the target, and the target of the same class have different instances need to be divided
Target Detection mainly to the position of the positioning target location information is generally represented as a rectangle, rectangle can be four-dimensional data to represent
object segmentation need to be divided into different categories for each pixel, the segmentation results need to be consistent with the original image size, often or by sampling the form of deconvolution to obtain an output result from the original image size