CN104268537A

CN104268537A - Main body detection method based on complex commodity images

Info

Publication number: CN104268537A
Application number: CN201410542974.2A
Authority: CN
Inventors: 彭学露; 陈永健; 黄琦
Original assignee: HANGZHOU TAOTAOSOU TECHNOLOGY Co Ltd
Current assignee: HANGZHOU TAOTAOSOU TECHNOLOGY Co Ltd
Priority date: 2014-10-14
Filing date: 2014-10-14
Publication date: 2015-01-07

Abstract

The invention discloses a main body detection method based on complex commodity images. Main body area detection is automatically conducted on the input complex images. The method includes the steps that human face detection is conducted, if a human face is detected, a watershed partition method based on the human face is adopted, and if the human face is not detected, a watershed partition method based on an image center area is adopted. The method is particularly suitable for being used in e-commerce and image search shopping websites, a background area is removed, a body area is emphasized, the requirement for the real-time performance of image retrieval is met, and the method serves for a shopping image search engine.

Description

A kind of based on complicated commodity image subject detection method

Technical field

The present invention relates to image procossing and image seek technology FIELD OF THE INVENTIONThe, specifically a kind of body region detection method for shopping items image (dress ornament, footwear and bag).

Background technology

Along with the development of ecommerce and shopping website, picture search receives much concern in academia and industry member.But commodity image scene is complicated, background is mixed and disorderly, multiple target co-occurrence, and illumination and object to block etc. and make picture search become a challenging difficult problem.At shopping website, most typical such as Taobao, seller is in order to attract client, and the modes such as the same money commodity spliced map of often selection street bat, Mo Tetu, multiple color represent commodity.The picture scene of complexity like this, brings very strong noise to picture search, and this can affect the effect of picture search greatly.For reducing the picture search influential effect that ground unrest brings, needing by automatically locating, accurately catching the body region of the conspicuousness in input commodity figure, effectively improving retrieval precision and the efficiency of " to scheme to search figure ", for follow-up feature extraction.

Early stage image subject method for detecting area, comprises Saliency maps method and the image region segmentation scheduling algorithm of computed image.Saliency maps calculates body region by calculating color contrast and spatial contrast degree, effectively can not process the close image of prospect background color.Image segmentation algorithm effectively can not process the mixed and disorderly image of background, and flase drop goes out object in background but not commodity body.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, provide a kind of based on complicated commodity image subject detection method.

The object of the invention is to be achieved through the following technical solutions.Comprise the steps: based on complicated commodity image subject detection method

(1) for the first size scaling of query image of user's input, Face datection is carried out;

(2) if detect face in step (1), face represents with rectangle frame, then take out the face that degree of confidence (with the matching degree of faceform) is maximum, adopts fractional spins, draws the estimation of commodity body rectangular area;

(3) if face do not detected in step (2), then default image central area is prospect, adopts the fractional spins based on picture centre, draws the estimation of commodity body;

The invention has the beneficial effects as follows, comprehensive human face detection tech and fractional spins carry out commodity image subject detection, improve accuracy rate, enhance robustness and the versatility of subject detection in complicated image.

Accompanying drawing explanation

Fig. 1 is the process flow diagram based on complicated commodity image subject detection method.

Embodiment

Here describes the present invention in detail by reference to the accompanying drawings, and object of the present invention and effect will become definitely.

Step 1, for the query image of user's input, on the basis keeping former figure the ratio of width to height, be that 310 pixels carry out convergent-divergent according to wide and high maximal value, the width of to be I, I.width the be image after convergent-divergent of the image after convergent-divergent, I.height is the height of image after convergent-divergent;

Step 2, Face datection is carried out to the query image of input, and export coordinate FC=Rect (FC.x, FC.y, the FC.width of human face region, FC.height), wherein Rect represents rectangle frame, and FC.x is the upper left corner x coordinate of FC, and FC.y is the upper left corner y coordinate of FC, FC.width is the width of FC, and FC.height is the height of FC;

Step 3, if detect face in step 2, if there is multiple face, only gets one that area is maximum; According to face information, image is divided into 3 subregions, foregrounding region F ₁=Rect (F ₁.x, F ₁.y, F ₁.width, F ₁.height), wherein Rect represents rectangle frame, F ₁.x be F ₁upper left corner x coordinate, F ₁.y be F ₁upper left corner y coordinate, F ₁.width be F ₁width, F ₁.height be F ₁height, background area comprises the image boundary region that pixel wide is S, and face upper area B ₁=Rect (B ₁.x, B ₁.y, B ₁.width, B ₁.height), wherein Rect represents rectangle frame, B ₁.x be B ₁upper left corner x coordinate, B ₁.y be B ₁upper left corner y coordinate, B ₁.width be B ₁width, B ₁.height be B ₁height, remaining area is uncertain region, for above 3 subregions, adopts Beucher S.The watershed transformation applied to image segmentation.Scanning Microsc 1992; The fractional spins of 6:299 – 314.d, then the foreground area after segmentation gets external square, is exactly the body region of commodity;

Background border area pixel width S calculating formula is

S＝0.1min(I.width,I.height)

Background area B ₁calculating formula be

B ₁.x＝0

B ₁.y＝0

B ₁.width＝I.width

B ₁.height＝FC.y+FC.height

Foreground area F ₁calculating formula be

F ₁.x＝max(0,FC.x+FC.width/3)

F ₁.y＝max(0,FC.y+1.5FC.width)

F ₁.width＝FC.width/2

F ₁.height＝3FC.height

Step 4, if face do not detected in step 2, then adopts the watershed algorithm based on picture centre, image is divided into 3 subregions, foregrounding region F ₂=Rect (F ₂.x, F ₂.y, F ₂.width, F ₂.height), wherein Rect represents rectangle frame, F ₂.x be F ₂upper left corner x coordinate, F ₂.y be F ₂upper left corner y coordinate, F ₂.width be F ₂width, F ₂.height be F ₂height, calculating formula is:

F ₂.x＝max(0,0.45I.width)

F ₂.y＝max(0,0.25I.height)

F ₂.width＝0.1I.width

F ₂.height＝0.5I.height

The image boundary region of to be pixel wide be in background area S, remaining area is uncertain region, S computing method are with step 3, for above 3 subregions, adopt Beucher S.The watershed transformation applied to image segmentation.Scanning Microsc 1992; The fractional spins of 6:299 – 314.d, the foreground area after segmentation gets external square, is exactly the body region of commodity.

Claims

1., based on a complicated commodity image subject detection method, it is characterized in that: comprise the steps:

(1) for the query image of user's input, size scaling is carried out, Face datection;

(2) if detect face in step (1), face represents with rectangle frame, then take out the face that degree of confidence (with the matching degree of faceform) is maximum, adopt the fractional spins based on face, draw the estimation of commodity body rectangular area;

(3) if face do not detected in step (2), adopt the fractional spins based on picture centre, draw the estimation of commodity body.