first, the main ideas
CNN has achieved very good results in the field of computer vision, but the model that trains CNN needs a lot of data, and small datasets don't show the benefits of CNN. The author proposes to solve this problem by transferring CNN weights . The first is to train the traditional CNN model on big data sets such as Imagenet, and then fine-tuning(possibly inaccurate) on specific tasks, but in the past fine-tuning The difference is: Change the whole pre-training The CNN framework has yielded good results on VOC2007 and VOC2012.
II. Basic Framework1. OverviewCNN hasmillionparameters, it is unrealistic to learn these parameters on a small data set, so it can be done on large scale datasetspre-training, and then use it on top of a specific task. such as:
However, there is a problem: the pre-trained data set and the specific task of the data set of the image is very different, such as the type of object, angle, image imaging conditions, and so on: In response to this question, the author proposes: (1) design a model that can be accurately remapped in a pre-training set and a specific task set(2) bysliding windowto improve the training and testing process
2. Network Framework(1) Pre-training with the traditional CNN framework on the Imagenet(2) Remove the last layer of Softmax layer, plus the FCA and FCB two layer adaptive layer(3) The parameters of the front 7 layer of the fixed pre-training model are unchanged, only the parameters of the adaptive layer are trained .3, network Training (preparation of training samples)(1) using the method of sliding window to extract 500 square image blocks per image, the coincidence ratio between each block is at least 50%;(2) to label the image block, assuming that the image block is P, a class of positive sample is Bo labeled as the corresponding class of positive sample conditions:(A) the intersection of P and Bo is 0.2 times times greater than or equal to the area of P(B)the intersection of P and Bo is 0.6 times times greater than the area of Bo (C) contains no more than one object in PExample:
(3) Working with background imagesusually the resulting sample will lead to the problem of training sample imbalance, most of the image block is a background image, (processing such a problem can use hard negative mining or re-change the weight of the loss function), this paper uses the random selection of background image 10%;4. ClassificationThe formula is as follows:
represents a class of scores in a single image. third, the experimentThe corresponding experiments were carried out on the VOC data set.
Conclusion: (1) The coincidence degree of the target category of the pre-training data set and the specific task data is not very big and small, but with the increase of the target type of the pre-training data set, the coincidence degree of the category increases, and the recognition rate is greatly improved. It is more important and impossible to know whether the training category has increased or the overlap has increased. (2) Increase or decrease the level of self-adaptation, the accuracy of 1% decline Iv. SummaryThe highlight of this paper is that in the fine-tuning of the time to remove the Softmax layer to increase the adaptive layer, as well as a statistical image of a certain category of scores, I think it is. However, one disadvantage is that when preparing a training sample, it may take a lot of time to manually determine how many objects are contained in the patch.
