Image feature computation and representation -- content based image retrieval

1️⃣ Job requirements

Given not less than 100 An appropriate set of images , They come in different sizes , Select any image , And a target area in the image is given manually , Like face 、 building 、 Dogs, etc , It is required to design a content-based image retrieval method , It can find... In the remaining images 5 Image containing the most similar framed object . It is better to roughly frame the target area of the retrieved image , No box is OK .

2️⃣ Core code

# coding:utf-8def extract_features(image_path, vector_size=32):    image = imageio.imread(image_path)    try:        # Using KAZE, cause SIFT, ORB and other was moved to additional module        alg = cv2.KAZE_create()        # Finding image keypoints        #  Find image keys         kps = alg.detect(image)        # Getting first 32 of them.        #  Before calculation 32 individual         # Number of keypoints is varies depend on image size and color pallet        #  The number of keys depends on the image size and the color palette         # Sorting them based on keypoint response value(bigger is better)        #  Sort according to the return value of the key （ The bigger the better ）        kps = sorted(kps, key=lambda x: -x.response)[:vector_size]        # computing descriptors vector        #  Compute descriptor vector         kps, dsc = alg.compute(image, kps)        # Flatten all of them in one big vector - our feature vector        #  Put it in a large vector , As our eigenvector         dsc = dsc.flatten()        # Making descriptor of same size        #  Make descriptors the same size         # Descriptor vector size is 64        #  The size of the descriptor vector is 64        needed_size = (vector_size * 64)        if dsc.size < needed_size:            # if we have less the 32 descriptors then just adding zeros            # at the end of our feature vector            #  If less than 32 A descriptor , Then zero is added after the eigenvector             dsc = np.concatenate([dsc, np.zeros(needed_size - dsc.size)])    except cv2.error as e:        print('Error: ', e)        return None    return dsc

3️⃣ experimental result

We randomly select a picture from the dataset ：

Frame an area in the image with the mouse ：

Then press enter to obtain the five pictures in the data set that are closest to the characteristics of the framed target area ：

The matching value of each picture will also be output , The matching value here is the cosine distance between the feature of the target area and the image in the database , If you want to be more similar , Data sets can be expanded , Make the features more suitable ：

Experiment source code + The report