1. Target detection algorithm based on sliding window
Sliding window>>feature extraction>>classifier
Figure sliding window target detection process
Set sliding windows of different sizes for the input image, determine the step size to traverse the entire image, perform feature extraction (SIFT, HOG, etc.) on the current selection box after each sliding, and use pre-trained classifiers (SVM, Adaboost, etc. ) to judge the probability that there is a target in the area. Get all the sliding windows that may have targets, because these windows will have parts with high repetitions, and finally use the non-maximum suppression (Non-Maximum Suppression, NMS) method to screen, and get the detection target after NMS screening.
Disadvantages
: The sliding window method is simple and easy to understand, but traversing images with windows of different sizes and ratios leads to redundant selection boxes; at the same time, the aspect ratio of the object needs to be considered when designing the window, which increases the design complexity; the features of manual design are less robust; The efficiency is low, and the sliding window method is not recommended for tasks with high real-time requirements.
2. The target detection algorithm based on Selective Search
uses information such as edges, textures, and colors in the image to ensure a high recall rate (Recall) in the case of selecting fewer (hundreds to thousands) windows. ).
Selective Search >> Feature Extraction >> Classifier
Merging of Graph Segmentation Regions Based on Similarity
First, use the image segmentation method to initialize the segmented area, then calculate the similarity between two adjacent areas, and merge the two areas that are most similar (judged by color, texture, size, and shape overlap) each time, and use search in each iteration The frame is positioned to merge the area until a complete area is left, and finally feature extraction and classification are performed.
Summary
The biggest disadvantage of the sliding window is that the selection frame is redundant, and the selective search can effectively remove the redundant candidate frame, which greatly reduces the amount of calculation.